We propose a data augmentation strategy to improve Automatic Target Recognition (ATR) from Infrared (IR) imagery using vision transformers. Our method leverages external IR image repositories to select relevant samples that can boost the diversity of the training data. By doing so, we improve the model ability to learn from the challenging regions of the training feature space. Our approach uses attention-based explanations to identify under-represented regions in the feature space of the training data. We leverage this information to search for new samples that complement the current training data by covering the sparse gaps in the feature space. We evaluate the proposed approach on a public dataset with IR imagery of multiple targets. We show that our method achieves a significant 2% improvement in ATR performance compared to a baseline model trained without augmentation. We also show that our method outperforms other data augmentation techniques that do not consider under-represented regions. These results demonstrate the effectiveness of our approach in an infrared scenario, where there is high intraclass variance and large training sets are expensive to obtain.

The limitations of a machine learning model can often be traced back to the existence of under-represented regions in the feature space of the training data. Data augmentation is a common technique that has been used to inflate training datasets with new samples to improve the model performance. However, these techniques usually focus on expanding the data in size and do not necessarily aim to cover the under-represented regions of the feature space. In this paper, we propose an Attention-guided Data Augmentation technique for Vision Transformers (ADA-ViT). Our framework exploits the attention mechanism in vision transformers to extract visual concepts related to misclassified samples. The retrieved concepts describe under-represented regions in the training dataset that contributed to the misclassifications. We lever-age this information to guide our data augmentation process by identifying new samples and using them to augment the training data. We hypothesize that this focused data augmentation populates under-represented regions and improves the model’s accuracy. We evaluate our framework on the CUB dataset and CUB-Families. Our experiments show that ADA-ViT outperforms state-of-the-art data augmentation strategies, and can improve the accuracy of a transformer by an average margin of 2.5% on the CUB dataset and 3.3% on CUB-Families.

We propose a Multiple Instance Learning (MIL) approach for lung nodules classification to address the limitations of current Computer-Aided Diagnosis (CAD) systems. One of these limitations consists of the need for a large collection of training samples that require to be segmented and annotated by radiologists. Another consists of using a fixed volume size for all nodules regardless of their actual sizes. Using a MIL approach, we represent each nodule by a nested sequence of volumes centered at the identified center of the nodule. We extract one feature vector from each volume. The set of features for each nodule are combined and represented by a bag. Using this representation, we investigate and compare many MIL algorithms and feature extraction methods. We start by applying benchmark MIL algorithms to traditional Gray Level Co-occurrence Matrix (GLCM) engineered features. Then, we design and train simple Convolutional Neural Networks (CNNs) to learn and extract features that characterize lung nodules. These extracted features are then fed to a benchmark MIL algorithm to learn a classification model. We report the results of three experiments applied to both GLCM and CNN features using two benchmark datasets. We designed our experiments to compare the different features and compare MIL versus Single Instance Learning (SIL) where a single feature vector represents a nodule. We show that our MIL representation using CNN features is more accurate for the lung nodules diagnosis task. We also show that MIL representation achieves better results than SIL applied on the ground truth region of each nodule.

Lung cancer is by far the leading cause of cancer death in the US. Recent studies have demonstrated the effectiveness of screening using low dose CT (LDCT) in reducing lung cancer related mortality. While lung nodules are detected with a high rate of sensitivity, this exam has a low specificity rate and it is still difficult to separate benign and malignant lesions. The ISBI 2018 Lung Nodule Malignancy Prediction Challenge, developed by a team from the Quantitative Imaging Network of the National Cancer Institute, was focused on the prediction of lung nodule malignancy from two sequential LDCT screening exams using automated (non-manual) algorithms. We curated a cohort of 100 subjects who participated in the National Lung Screening Trial and had established pathological diagnoses. Data from 30 subjects were randomly selected for training and the remaining was used for testing. Participants were evaluated based on the area under the receiver operating characteristic curve (AUC) of nodule-wise malignancy scores generated by their algorithms on the test set. The challenge had 17 participants, with 11 teams submitting reports with method description, mandated by the challenge rules. Participants used quantitative methods, resulting in a reporting test AUC ranging from 0.698 to 0.913. The top five contestants used deep learning approaches, reporting an AUC between 0.87 - 0.91. The team's predictor did not achieve significant differences from each other nor from a volume change estimate (p =.05 with Bonferroni-Holm's correction).

We propose discrimination algorithms for buried threat detection (BTD) that exploit deep convolutional neural networks (CNNs) and recurrent neural networks (RNN) to analyze 2-D GPR B-scans in the down-track (DT) and cross-track (CT) directions as well as 3-D GPR volumes. Instead of imposing a specific model or handcrafted features, as in most existing detectors, we use large real GPR data collections and data-driven approaches that learn: 1) features characterizing buried explosive objects (BEOs) in 2-D B-scans, both in the DT and CT directions; 2) the variation of the CNN features learned in a fixed 2-D view across the third dimension; and 3) features characterizing BEOs in the original 3-D space. The proposed algorithms were trained and evaluated using large experimental GPR data covering a surface area of 120 000 m2from 13 different lanes across two U.S. test sites. These data include a diverse set of BEOs consisting of varying shapes, metal content, and underground burial depths. We provide some qualitative analysis of the proposed algorithms by visually comparing their performance and consistency along different dimensions and visualizing typical features learned by some nodes of the network. We also provide quantitative analysis that compares the receiver operating characteristics (ROCs) obtained using the proposed algorithms with those obtained using existing approaches based on CNN as well as traditional learning.

In this paper, we consider the development of algorithms for the automatic detection of buried threats using ground penetrating radar (GPR) measurements. GPR is one of the most studied and successful modalities for automatic buried threat detection (BTD), and a large variety of BTD algorithms have been proposed for it. Despite this, large-scale comparisons of GPR-based BTD algorithms are rare in the literature. In this paper, we report the results of a multi-institutional effort to develop advanced BTD algorithms for a real-world GPR BTD system. The effort involved five institutions with substantial experience with the development of GPR-based BTD algorithms. In this paper, we report the technical details of the advanced algorithms submitted by each institution, representing their latest technical advances, and many state-of-the-art GPR-based BTD algorithms. We also report the results of evaluating the algorithms from each institution on the large experimental data set used for development. The experimental data set comprised 120 000 m2of GPR data using surface area, from 13 different lanes across two U.S. test sites. The data were collected using a vehicle-mounted GPR system, the variants of which have supplied data for numerous publications. Using these results, we identify the most successful and common processing strategies among the submitted algorithms, and make recommendations for GPR-based BTD algorithm design.

We present a framework to detect and retrieve CPR activity scenes from medical simulation videos. Medical simulation is a modern training method for medical practitioners, where an emergency patient condition is simulated on humanlike mannequins and the students act upon. These simulation sessions are recorded for later debriefing. With the increasing number of simulation videos, automatic detection and retrieval of specific scenes became necessary. In this paper, we present an automated CPR scene retrieval system which can identify and retrieve CPR activity scenes, using Local Binary Patterns Over Three Orthogonal Planes (LBP-TOP) features. We present a comparative study of LBP-TOP features and Histogram of Orientation of Gradients (HOG3D) features for CPR action detection. We also present the results of decision level fusion of different classifier outputs. Promising classification results have been achieved, suggesting the proposed technique to be effective in detecting and retrieving CPR activity scenes from simulation videos.

Multiple instance regression (MIR) operates on a collection of bags, where each bag contains multiple instances sharing an identical real-valued label. Only few instances, called primary instances, contribute to the bag's label. The remaining instances are noise and outliers observations. The goal in MIR is to identify the primary instances within each bag and learn a regression model that can predict the label of a previously unseen bag. In this paper, we present an algorithm that uses robust fuzzy clustering with an appropriate distance to learn multiple linear models from a noisy feature space simultaneously. We show that fuzzy memberships are useful in allowing instances to belong to multiple models, while possibilistic memberships allow identification of the primary instances of each bag with respect to each model. We also use the possibilistic memberships to identify the optimal number of regression models. We evaluate our approach on a series of synthetic data sets. We show that our approach achieves higher accuracy than existing methods.

SCoTS captures a sparse representation of shapes in an input image through a linear span of previously delineated shapes in a training repository. The model updates shape prior over level set iterations and captures variabilities in shapes by a sparse combination of the training data. The level set evolution is therefore driven by a data term as well as a term capturing valid prior shapes. During evolution, the shape prior influence is adjusted based on shape reconstruction, with the assigned weight determined from the degree of sparsity of the representation. For the problem of lung nodule segmentation in X-ray CT, SCoTS offers a unified framework, capable of segmenting nodules of all types. Experimental validations are demonstrated on 542 3-D lung nodule images from the LIDC-IDRI database. Despite its generality, SCoTS is competitive with domain specific state of the art methods for lung nodule segmentation.

Fuzzy logic is a powerful tool to model knowledge uncertainty, measurements imprecision, and vagueness. However, there is another type of vagueness that arises when data have multiple forms of expression. This is the case for multiple instance learning problems (MIL). In MIL, an object is represented by a collection of instances, called a bag. Labels of bags are known but not those of individual instances. In previous work, we have introduced MI-ANFIS, a Multiple Instance-ANFIS, that extends the standard Adaptive Neuro Fuzzy Inference System (ANFIS) to handle reasoning with bags of instances as input and capable of learning from ambiguously labeled data. In this paper, first, we present a regularization technique, we called Rule Dropout, and show that it could be used to train MI-ANFIS systems with better generalization. Next, we develop a multi class MI-ANFIS (MCMI-ANFIS), that could be used to solve multiple class classification problems effectively. The proposed MCMI-ANFIS is tested and validated using a benchmark data set suitable for multiple class MIL problems.

We introduce a new algorithm that maps multiple instance data using both positive and negative target concepts into a data representation suitable for standard classification. Multiple instance data are characterized by bags which are in turn characterized by a variable number of feature vectors or instances. Each bag has a known positive or negative label, but the labels of any given instances within a bag is unknown. First, we use the Fuzzy Clustering of Multiple Instance data (FCMI) algorithm to identify K+positive target concepts, which represent points in the feature space that are close to instances from positive bags, and distant to instances from negative bags. We use a simple K-means clustering algorithm to identify K-negative target concepts that supplement the positive target concepts. Next we demonstrate how the positive and negative target concepts can be used to embed each bag, which has a variable number of instances, into a feature vector with fixed dimension. A key advantage to embedded instance space feature vectors is that standard machine learning algorithms may be used in training and testing multiple instance data. Another advantage of our embedding is that it provides a simple and intuitive interpretation of the data. We show that using our feature embedding, coupled with standard classifiers such as support vector machines or k-nearest neighbors, can outperform state-of-the-art Multiple Instance Learning classifiers on benchmark datasets.

We introduce a new algorithm to identify multiple target concepts when data are represented by multiple instances. A multiple instance data sample is characterized by a bag that contains multiple feature vectors, or instances. Each bag is labeled as either positive or negative. However, the labels of the instances within each bag are unknown. A bag is labeled as positive if and only if at least one of its instances is positive and negative if and only if all of its instances are negative. First, we define a fuzzy Multi-target concept Diverse Density (MDD) metric. The MDD is maximized when the target concepts correspond to dense regions in the feature space with maximal correlation to instances from positive samples, and minimal correlation to instances from negative samples. Then, we develop an iterative algorithm to optimize the MDD and identify K target concepts simultaneously. The proposed algorithm, called Fuzzy Clustering of Multiple Instance data (FCMI), is tested and validated by using it to analyze data of buried landmines collected using a ground penetrating radar sensor. We show that the FCMI algorithm can identify distinct target concepts that correspond to mines of different types buried at different depths. We also show that FCMI can be used to label individual instances within each bag.

We introduce a novel adaptive neuro-fuzzy architecture based on the framework of Multiple Instance Fuzzy Inference. The new architecture called Multiple Instance-ANFIS (MI-ANFIS), is an extension of the standard Adaptive Neuro Fuzzy Inference System (ANFIS) [1] that is designed to handle reasoning with multiple instances (bags of instances) as input and capable of learning from ambiguously labeled data. In multiple instance problems the training data is ambiguously labeled. Instances are grouped into bags, labels of bags are known but not those of individual instances. Multiple Instance Learning (MIL) deals with learning a classifier at the bag level. Over the years many solutions to this problem have been proposed. However, no MIL formulation employing fuzzy inference exists in the literature. In this paper, we develop MI-ANIFS that generalizes ANFIS inference systems to account for ambiguity and reason with multiple instances. We also develop a learning algorithm to learn the parameters of MI-ANFIS. The proposed MI-ANFIS is tested and validated using a synthetic and benchmark data sets suitable for MIL problems.

We present a robust speaker identification algorithm that uses effective features based on Fisher Vector (FV) representations. First, low-level spectral features are extracted from the training data. Next, we model the data (in the spectral feature space) by a mixture of Gaussian components. Then, we construct FV descriptors based on the deviation of the features from the Gaussian components. We analyze the FV feature representations on speech data with two common classifiers: K-nearest neighbor classifier (KNN) and support vector machines (SVM). The proposed approach is evaluated using audio data sets recorded to simulate medical crises. We show that the proposed FV feature representation approach achieves a significant improvement when compared to the state-of-art methods.

We present a novel method to fuse the output of multiple discrimination algorithms for the purpose of landmine detection using Ground Penetrating Radar (GPR). The proposed fusion method is based on a neuro-fuzzy architecture employing Multiple Instance Fuzzy Inference, called Multiple Instance Adaptive Neuro Fuzzy Inference System (MI-ANFIS). Multiple Instance Fuzzy Inference is a generalization of fuzzy inference that enables fuzzy reasoning with bags of instances. Pre-processed GPR data are usually grouped into bags of signal slices extracted at multiple depths. Labels of the bags are known, but not those of individual instances. It is very difficult to localize the objects depth automatically, and it is a very tedious process to do it manually. MI-ANFIS is capable of learning meaningful and simple fusion rules from ambiguously labeled data. Thus, making it suitable for the purpose of multiple landmine discrimination algorithms fusion. Initial testing on large and diverse GPR data collections has shown promising results. MI-ANFIS was able to overcome labeling ambiguity and outperformed other commonly used fusion methods.

A new class of subpixel target detection algorithms that use a local structured background model is introduced. Our approach, referred to as Context Dependent Target Detectors, extends existing structured detectors to multiple contexts. It is based on a robust context dependent spectral unmixing algorithm that uses multiple linear models to take into account the background variability. The claim is that robust context dependent unmixing provides a better description of the background with minimum target leakage, compared to global unmixing, and hence results in a better target-background separation. The approach is evaluated using the Orthogonal Subspace Projection (OSP) detector, the Adaptive Matched Subspace Detector (AMSD) and the Hybrid Subspace Detector (HSD). Experimental results show that our approach outperforms the baseline detectors (that use one global model for background).

A robust hyper spectral unmixing algorithm that finds multiple sets of end members is introduced. The algorithm, called Robust Context Dependent Spectral Unmixing (RCDSU), combines the advantages of context dependent unmixing and robust clustering. RCDSU adapts the unmixing to different regions, or contexts, of the spectral space. It combines fuzzy and possibilistic clustering and linear unmixing to learn multiple contexts and the optimal end members and abundances for each context. RCDSU uses fuzzy membership functions to partition the data, and possibilistic membership functions to identify noise and outliers. An extension of RCDSU to deal with the case of an unknown number of contexts is also proposed. The performance of the proposed work is evaluated using simulated and real hyper spectral data. The experiments show that the proposed methods can handle noisy data and identify an "optimal" number of contexts and appropriate end members within each context.

We present a fusion method, based on fuzzy inference, for detecting buried objects using ground-penetrating radar (GPR) data. The performance of different discrimination algorithms can vary significantly depending on the target type, burial orientation, and other environmental conditions. In some cases, algorithms can provide complementary evidence, while in other cases they can provide contradicting evidence. Thus, effective fusion of these algorithms can achieve higher probability of detection with fewer false alarms. The proposed fusion method is based on an Adaptive Neuro Fuzzy Inference System (ANFIS) [1] capable of simultaneously identifying local contexts as well as learning optimal weights for combining local expert discriminators. It is capable of learning meaningful and simple fuzzy rules for different regions of the input space. Results on large and diverse GPR data collections show that the proposed fusion approach can identify local, simple, and meaningful rules capable of non-linear fusion of different discriminators. We also show that the proposed fuzzy inference outperforms other commonly used fusion methods.

A hyperspectral endmember detection and spectral unmixing algorithm that finds multiple sets of endmembers is presented. Hyperspectral data are often nonconvex. The Piecewise Convex Multiple-Model Endmember Detection algorithm accounts for this using a piecewise convex model. Multiple sets of endmembers and abundances are found using an iterative fuzzy clustering and spectral unmixing method. The results indicate that the piecewise convex representation estimates endmembers that better represent hyperspectral imagery composed of multiple regions where each region is represented with a distinct set of endmembers.

While classical kernel-based clustering algorithms are based on a single kernel, in practice it is often desirable to base clustering on combination of multiple kernels. In [1], we considered a fuzzy c-means with multiple kernels in observation space (FCMK-OS) algorithm which constructs the kernel from a number of Gaussian kernels and learns a resolution specific weight for each kernel function in each cluster. The FCMK-OS did not have a closed form expression to update the kernel weights. Moreover, it derives the fuzzy c-means in input space with kernelization of the metric. Thus, it can not handle nonlinear partitioning of the data. In this paper, we propose a fuzzy c-means with multiple kernels in feature space (FCMK-FS) algorithm which extends the fuzzy c-means algorithm with an adaptive multiple kernel learning setting. The incorporation of multiple kernels and unsupervised adjusting of the kernel weights in each cluster makes the choice of the kernels less crucial and allows better characterization and adaptability to each individual cluster. Experiments on both toy and real data sets demonstrate the effectiveness of the proposed FCMK-FS algorithm.

A hyperspectral endmember detection and spectral unmixing algorithm that finds multiple sets of endmembers is presented. This algorithm, the Piece-wise Convex Multiple Model Endmember Detection (P-COMMEND) algorithm, models a hyperspectral image using a piece-wise convex representation. By using a piece-wise convex representation, non-convex hyperspectral data are more accurately characterized. For example, the well-known Indian Pines hyperspectral image is used as an example of a piece-wise convex collection of pixels. The convex regions, weights, endmembers and abundances are found using an iterative fuzzy clustering method. Results indicate that the piece-wise convex representation provides endmembers that better represent hyperspectral data sets over methods that use a single convex region.

We propose a feature level fusion that is based on mapping the original low-level audio features to histogram descriptors. Our mapping is based on possibilistic membership functions and has two main components. The first one consists of clustering each set of features and identifying a set of representative prototypes. The second component uses the learned prototypes within membership functions to transform the original features into histograms. The mapping transforms features of different dimensions to histograms of fixed dimensions. This makes the fusion of multiple features less biased by the dimensionality and distributions of the different features. Using a standard collection of songs, we show that the transformed features provide higher classification accuracy than the original features. We also show that mapping simple low-level features and using a K-NN classifier provides results comparable to the state-of-the art.

We propose a machine learning based speaker segmentation and identification system that provides the physician with automated tools to segment, semantically index and retrieve specific segments from a large database of medical simulation video sessions. Instead of working directly in the original feature space, our approach maps low-level audio features to more meaningful histogram descriptors using possibilistic membership functions. The parameters of the membership functions are learned from representative training data. Using 4 medical simulation videos, we show that our approach outperforms existing speaker identification algorithms using standard MFCC and PLP speech features. The proposed speaker identification algorithm was integrated in a GUI to facilitate video retrieval and analysis. Using this GUI, the physician can efficiently identify who spoke and when. In addition, our system can extract useful statistics and features in a completely unsupervised way.

We propose a novel method for fusing different classifiers outputs. Our approach, called context extraction for local fusion with fuzzy integrals (CELF-FI), is a local approach that adapts a fuzzy integrals fusion method to different regions of the feature space. It is based on a novel objective function that combines context identification and multialgorithm fusion criteria into a joint objective function. This objective function consists of two terms: The first is designed to produce compact clusters, called contexts, and the second is designed to produce Sugeno measures for fuzzy integral fusion for each context. The terms are optimized simultaneously via alternating optimization. To test a new sample, first, its features (extracted by each algorithm) are used to assign it to each context with a fuzzy membership degree. Second, the sample confidence values (assigned by each algorithm) are fused within each context using the learned context fusion parameters. Then, the context-dependent partial confidence values are weighted by the membership degrees and averaged over all contexts to produce a final confidence value. We illustrate the performance of CELF using synthetic data, and we apply it to the problem of landmine detection using ground penetrating radar and wideband electromagnetic induction. Our extensive experiments have indicated that the proposed fusion approach outperforms all individual classifiers, the global fuzzy integral fusion method, and the basic local fusion with linear aggregation.

In this paper, the kernel fuzzy c-means clustering algorithm is extended to an adaptive cluster model which maps data points to a high dimensional feature space through an optimal convex combination of homogenous kernels with respect to each cluster. This generalized model, called Fuzzy C Means with Multiple Kernels (FCM-MK), strives to find a good partitioning of the data into meaningful clusters and the optimal kernel-induced feature map in a completely unsupervised way. It constructs the kernel from a number of Gaussian kernels and learns a resolution specific weight for each kernel function in each cluster. This allows better characterization and adaptability to each individual cluster. The effectiveness of the proposed algorithm is demonstrated for several toy and real data sets.

We propose a new relational clustering approach, called Fuzzy clustering with Learnable Cluster dependent Kernels (FLeCK), that learns multiple kernels while seeking compact clusters. A Gaussian kernel is learned with respect to each cluster. It reflects the relative density, size, and position of the cluster with respect to the other clusters. These kernels are learned by optimizing both the intra-cluster and the inter cluster similarities. Moreover, FLeCK is formulated to work on relational data. This makes it applicable to data where objects cannot be represented by vectors or when clusters of similar objects cannot be represented efficiently by a single prototype. The experiments show that FLeCK outperforms several other algorithms. In particular, we show that when data include clusters with various inter and intra cluster distances, learning cluster dependent kernel is crucial in obtaining a good partition.

We propose a novel image database categorization approach using robust unsupervised learning of finite generalized dirichlet mixture models with feature discrimination. The proposed algorithm is based on optimizing an objective function that associates two types of memberships with each data sample. The first one is the posterior probability and indicates how well a sample fits each estimated distribution. The second membership represents the degree of typicality and is used to identify and discard noise points and outliers. In addition, RULe_GDM learns an optimal relevance weight for each feature subset within each cluster. These properties make RULe_GDM suitable for noisy and high-dimensional feature spaces. We also extend our algorithm to find the optimal number of clusters in an unsupervised and efficient way by exploiting some properties of the possibilistic membership function. RULe_GDM is used to categorize a collection of color images. The performance of RULe_GDM is illustrated and compared to similar algorithms.

We propose a landmine detection algorithm using ground penetrating radar (GPR) data that uses multiple features and an ensemble of continuous hidden Markov models (CHMMs). Our approach is motivated by the need for different features and multiple models to capture the large variations of mine and clutter types and the different environmental conditions. First, for each feature representation, we fit one continuous HMM to each of the N alarm signatures. For each fitted model, we evaluate the log-likelihood of each sequence. This will result in one N × N log-likelihood distance matrix for each feature. These log-likelihood matrices are combined and partitioned to identify K clusters of alarms where each cluster contain alarms that are similar with respect to one or more features. Second, we learn the parameters of one HMM per group using the relevant features identified in the clustering phase. The mixture of all learned models would be used to build a descriptive model of the data. The third step of our approach consists of training a Neural Network (NN) to fuse the results of the mixture models and assign a final confidence value. Results on large and diverse Ground Penetrating Radar data collections show that the proposed approach can identify meaningful and coherent HMM models that describe different properties of the data. Our initial experiments have also indicated that the proposed mixture model outperforms the baseline mixture model that uses one single feature representation.

The Edge Histogram Detector (EHD) is a well-researched and tested algorithm that has been integrated into fielded systems for landmine detection using Ground Penetrating Radar (GPR) sensor. It uses edge histogram based features and a possibilistic K-Nearest Neighbors (KNN) classifier. Due to the inherent static data representation and static classifier architecture, the EHD may not be very effective in detecting targets with large variations in shape and size. In this paper, we propose a more flexible approach that is based on multiple instance learning. First, we summarize the training data and identify representative mine alarms. This summarization step could be achieved using one or multiple feature representation to capture different characteristics of the data. The identified prototypes, also called bags of mines, will be used to map the alarms to a feature space that improves the discrimination between mines and clutter objects. The second step of our approach consists of building a classifier on the mapped feature space. We experiment with two different classifiers. The first one is a simple linear classifier that compares the features in the mapped space. The second classifier is based on learning Relevance Vector Machines (RVM) in the sparse mapped space. Our initial experiments on large and diverse Ground Penetrating Radar data collections show that the proposed approach can outperform the baseline EHD.

We propose a multistream discrete hidden Markov model (DHMM) framework and apply it to the problem of land-mine detection using ground-penetrating radar (GPR). We hypothesize that each signature (mine or nonmine) can be characterized better by multiple synchronous sequences representing features that capture different environments and different radar characteristics. This paper is motivated by the fact that mines and clutter objects can have different characteristics depending on the mine type, soil and weather conditions, and burial depth. Thus, ideally different sets of specialized feature extraction mechanisms may be needed to achieve high detection and low false alarm rates. In order to fuse the different modalities, a multistream DHMM that includes a stream relevance weighting component is developed. The relevance weight of each stream depends on the symbols and the states. We reformulate the Baum-Welch and the minimum classification error/gradient probabilistic descent learning algorithms to include stream relevance weights and partial state probabilities. We generalize their objective functions and derive the necessary conditions to update all model parameters simultaneously. The results on a synthetic data set and a collection of GPR signatures show that the proposed multistream DHMM framework outperforms the basic single-stream DHMM where all the streams are treated equally important.

We propose a novel method for fusing different classifiers outputs. Our approach, called Context Extraction for Local Fusion with Fuzzy Integrals (CELF-FI), is a local approach that adapts fuzzy integrals fusion method to different regions of the feature space. It is based on a novel objective function that combines context identification and multi-algorithm fusion criteria into a joint objective function. This objective function is defined and optimized to produce contexts as compact clusters via unsupervised clustering. Optimization of the objective function also provide an optimal Sugeno measure within each context. Our initial experiments have indicated that the proposed fusion approach outperforms all individual classifiers, the global fuzzy integral fusion method, and the basic local fusion with linear aggregation.

Context extraction for local fusion (CELF) is a local approach that combines multiple classifier outputs with the help of feature space information. CELF is based on an objective function that integrates context extraction and decision fusion. Context extraction divides the feature space into homogeneous regions; decision fusion combines multiple classifier outputs in each region or context. Although CELF is a generalization of previous fusion methods, it has some problems and noise sensitivity is one of them. The noise sensitivity problem comes from the fact that there are large number of parameters that should be optimized. Some variants of CELF were propose to mitigate the noise sensitivity problem and they showed better results than CELF. However, they also have their own shortcomings. In this paper, we proposed another variant of CELF to reduce the noise sensitivity and to overcome the shortcomings in the variants of CELF. The proposed method, CELF with regularization (CELF-R), uses regularization, which is a well-known method to reduce noise sensitivity by smoothing solution space. The noise robustness in CELF-R helps it to achieve better results than CELF and its variants. Experiments using landmine data sets also support this.

We propose a novel possibilistic clustering algorithm based on robust modelling of the Generalized Dirichlet (GD) finite mixture. The algorithm generates two types of membership degrees. The first one is a posterior probability that indicates the degree to which the point fits the estimated distribution. The second membership represents the degree of “typicality” and is used to indentify and discard noise points. The algorithm minimizes one objective function to optimize GD mixture parameters and possibilistic membership values. This optimization is done iteratively by dynamically updating the Dirichlet mixture parameters and the membership values in each iteration. We compare the performance of the proposed algorithm with an EM based approach. We show that the possibilistic approach is more robust.

We propose a landmine detection algorithm using ground penetrating radar data that is based on an SVM classifier. The kernel function for the SVM is constructed using discrete hidden Markov modeling (HMM). Typically, the kernel matrix could be obtained by defining an adequate similarity measure in the feature space. However, this approach is inappropriate as it is not trivial to define a meaningful distance metric for sequence comparison. Our proposed approach is based on HMM modeling and has two main steps. First, one HMM is fit to each of the N individual sequences. For each fitted model, we evaluate the log-likelihood of each sequence. This will result in an N × N log-likelihood similarity matrix that will be adapted to serve as the kernel of the SVM classifier. In the second step, we train an SVM classifier to learn a decision boundary between the positive and negative samples.

We propose a discriminative method for combining heterogeneous sets of features for the continuous hidden Markov model classifier. We use a model level fusion approach and apply it to the problem of landmine detection using ground penetrating radar (GPR). We hypothesize that each signature (mine or non-mine) can be characterized better by multiple synchronous sequences that can capture different and complementary features. Our work is motivated by the fact that mines and clutter objects can have different characteristics depending on the mine type, soil and weather conditions, and burial depth. Thus, different sets of specialized feature extraction mechanisms, may be needed to achieve high detection and low false alarm rates. In order to fuse the different modalities, a multi-stream continuous HMM that includes a stream relevance weighting component is developed. In particular, we modify the probability density function that characterizes the standard continuous HMM to include state and component dependent stream relevance weights. We generalize the Minimum Classification Error (MCE) objective function to include stream relevance weights and derive the necessary conditions to update all model parameters simultaneously. Results on a large collection of GPR alarms show that the proposed model level fusion outperforms the baseline HMM when each feature is used independently and when both features are combined with equal weights.

We present a novel method for fusing the results of multiple land mine detection algorithms which use different sensors, features, and different classification methods. The proposed multisensor/multialgorithm fusion method, which is called context-dependent fusion (CDF), is motivated by the fact that the relative performance of different sensors and algorithms can vary significantly depending on the mine type, geographical site, soil and weather conditions, and burial depth. CDF is a local approach that adapts the fusion method to different regions of the feature space. The training part of CDF has two components: context extraction and algorithm fusion. In context extraction, the features used by the different algorithms are combined and used to partition the feature space into groups of similar signatures, or contexts. The algorithm fusion component assigns a degree of worthiness to each detector in each context based on its relative performance within the context. To test a new alarm using CDF, each detection algorithm extracts its set of features and assigns a confidence value. Then, the features are used to identify the best context, and the degrees of worthiness of this context are used to fuse the individual confidence values. Results on large and diverse ground-penetrating radar and wideband electromagnetic data collections show that the proposed method can identify meaningful and coherent clusters and that different expert algorithms can be identified for the different contexts. Typically, the contexts correspond to groups of alarm signatures that share a subset of common features. Our extensive experiments have also indicated that CDF outperforms all individual detectors and the global fusion that uses the same method to assign aggregation weights.

We present a novel method for fusing different classifiers outputs. Our approach, called Context Extraction for Local Fusion with Feature Discrimination (CELF-FD), is a local approach that adapts the fusion method to different regions of the feature space. It is based on a novel objective function that combines context identification and multi-algorithm fusion criteria into a joint objective function. This objective function is defined and optimized to produce contexts as compact clusters in subspaces of the high-dimensional feature space via unsuper-vised clustering and feature discrimination. Optimization of the objective function also provide optimal fusion parameters for each context. Our initial experiments have indicated that the proposed fusion approach outperforms all individual classifiers and the global fusion method.

In this paper, we propose a new general low-level feature representation for audio signals. Our approach, called Dominant Audio Descriptor is inspired by the MPEG-7 Dominant Color Descriptor. It is based on clustering time-local features and identifying dominant components. The features used to illustrate this approach are the well-known Mel Frequency Cepstral Coefficients. The performance of the proposed framework is evaluated on audio classification and retrieval tasks. In particular, the experiments are performed on a benchmark music data set. The results are compared to those previously obtained on the same data base. We show that our approach improved classification and retrieval results by more then 3%, and for the case of retrieval reached almost perfect retrieval rate of 99:36%. In addition, the paper presents comparative results against several state of the art classifiers, such as Hidden Markov Models, Support Vector Machines and k-Nearest Neighbors.

This paper describes the Ensemble Possibilistic K-NN algorithm for classification of gene expression profiles into three major cancer categories. In fact, a modification of forward feature selection is proposed to identify relevant feature subsets allowing for multiple possibilistic K-nearest neighbors (pK-NNs) rule experts. First, individual features are ranked according to their performance on training data and subsets of features identified using greedy approach. Each subset has significantly lower dimensionality than the original feature vector. Second, each subset is associated with pK-NN expert and the final classification decision is based on combining results produced by all experts.

We propose a modified discrete HMM that handles multimodalities. We assume that the feature space is partitioned into subspaces generated by different sources of information. To combine these heteregoneous modalities we propose a multi-stream discrete HMM that assigns a relevance weight to each subspace. The relevance weights are set local and depend on the symbols and the states. In particular, we associate a partial probability with each symbol in each subspace. The overall observation state probability is then computed as an aggregation of the partial probabilities and their objective relevance weights based on a linear combination. The minimum classification error (MCE) objective based on the gradient probabilistic descent (GPD) optimization algorithm is reformulated to derive the update equations for the relevance weights and the partial state probabilities. The proposed approach is validated using synthetic and real data sets. The results are shown to outperform the baseline discrete HMM that treats all streams equally important.

We present a novel method for fusing the decisions of multiple classification algorithms which use different features, classification methods, and data sources. The proposed method, called context dependent fusion of multiple algorithms (CDF-MA) is motivated by the fact that the relative performance of different algorithms can vary significantly as the characteristics of the input data vary. The training part of CDF-MA has two main components: context extraction and algorithm fusion. In context extraction, the features used by the distinct algorithms are combined and clustered into groups of similar contexts. The algorithm fusion component embeds a variation of the MCE/GPD method to assigns an aggregation weight to each algorithm based on its loss function within each context. Results on real world data show that the proposed method can identify meaningful and coherent clusters, and outperform all individual classifiers and the global weighted average fusion method.

Principal component analysis (PCA) is widely used for dimensionality reduction in pattern recognition. Although PCA has been applied in many areas successfully, it suffers from sensitivity to noise and is limited to linear principal components. The noise sensitivity problem comes from the least-squares measure used in PCA and the limitation to linear components originates from the fact that PCA uses an affine transform defined by eigenvectors of the covariance matrix and the mean of the data. In this paper, a robust kernel PCA method that extends Scholkopf et al.'s kernel PCA and uses fuzzy memberships is introduced to tackle the two problems simultaneously. We first propose an iterative method to find a robust covariance matrix called robust fuzzy PCA (RF-PCA). The RF-PCA is introduced to reduce the sensitivity to noise with the help of robust estimation technique. The RF-PCA method is then extended to a non-linear one, robust kernel fuzzy PCA (RKF-PCA), using kernels. Experimental results suggest that the proposed algorithm works well on artificial and real world data sets.

We propose an image annotation approach that relies on fuzzy clustering and feature discrimination, a greedy selection and joining algorithm (GSJ), and Bayes rule. Clustering is used to group image regions into prototypical region clusters that summarize the training data and can be used as the basis of annotating new test images. Since this problem involves clustering sparse and high dimensional data, we use a semi-supervised constrained clustering algorithm that performs simultaneous clustering and feature discrimination. The constraints consist of pairs of image regions that should not be included in the same cluster. These constraints are deduced from the irrelevance of all concepts annotating the training images. The constraints help in guiding the clustering process. The GSJ algorithm uses the fuzzy membership function of each region cluster. Finally, Bayes rule is used to label images based on the posterior probability of each concept. The proposed algorithm was implemented and tested on a data set that includes 3000 images using four-fold cross validation.

This paper describes an algorithm for land mine detection using sensor data generated by a ground-penetrating radar (GPR) system that uses edge histogram descriptors for feature extraction and a possibilisticK-nearest neighbors (K-NNs) rule for confidence assignment. The algorithm demonstrated the best performance among several high-performance algorithms in extensive testing on a large real-world datasets associated with the difficult problem of land mine detection. The superior performance of the algorithm is attributed to the use of the possibilisticK-NN algorithm, thereby providing important evidence supporting the use of possibilistic methods in real-world applications. The GPR produces a 3-D array of intensity values, representing a volume below the surface of the ground. First, a computationally inexpensive prescreening algorithm for anomaly detection is used to focus attention and identify candidate signatures that resemble mines. The identified regions of interest are processed further by a feature extraction algorithm to capture their salient features. We use translation-invariant features that are based on the local edge distribution of the 3-D GPR signatures. Specifically, each 3-D signature is divided into subsignatures, and the local edge distribution for each subsignature is represented by a histogram. Next, the training signatures are clustered to identify prototypes. The main idea is to identify few prototypes that can capture the variations of the signatures within each class. These variations could be due to different mine types, different soil conditions, different weather conditions, etc. Fuzzy memberships are assigned to these representatives to capture their degree of sharing among the mines and false alarm classes. Finally, a possibilisticK-NN-based rule is used to assign a confidence value to distinguish true detections from false alarms. The proposed algorithm is implemented and integrated within a complete land mine prototype system. It is trained, field-tested, evaluated, and compared using a large-scale cross-validation experiment that uses a diverse dataset acquired from four outdoor test sites at different geographic locations. This collection covers over 41 807 m2of ground and includes 1593 mine encounters.

The optimal combination of the outputs of multi-modal features in content-based image retrieval (CBIR) is an important task that can have a significant impact on the overall performance of the CBIR system. This problem has not received much attention from the CBIR research community, and only simple methods have been used. In this paper, we treat the problem as an information fusion problem and propose an approach that is generic and can be adapted to various features and distance measures using a small set of training images. Our approach is based on associating a fuzzy membership function with the distribution of the features’ distances, and assigning a degree of worthiness to each feature based on its average performance. The memberships and the feature weights are then aggregated to produce a confidence that could be used to rank the retrieved images. We describe and experiment with two distinct aggregation methods. The first one is linear and is based on a simple weighted combination. The second one is non-linear and is based on the discrete Choquet integral. The proposed fusion methods were trained and tested using a large collection of images with several low-level visual and high-level textual features. The results were compared to other methods used in typical CBIR systems.

We propose an adaptive constrained clustering (ACC) algorithm that performs clustering and feature weighting simultaneously and that can incorporate partial supervision information. This information consists of a set of constraints on which instances should or should not reside in the same cluster. The algorithm is dynamic in the sense that the optimal number of clusters can expand or shrink depending on the distribution of the data and the set of constraints. The ACC algorithm is used as the main tool to organize and navigate through a collection of geo-referenced images. In this application, the constraints are generated automatically based on the spatial distribution of the coordinates of the images and the dynamic range of the current view. As a result, images would be assigned to the same cluster if they are similar in content and close spatially. Icons of the clusterspsila representatives are superimposed on a map to provide the user with a global overview of the content and location of the photo collection. If the user selects a region to zoom-in, images within this region will be re-clustered with stricter spatial constraints. The proposed clustering and visualization application were applied to a collection of 2,023 photos taken with a digital camera that tags each photo with the location of where the photo was taken using a GPS receiver.

In this paper, we propose a generic approach for representing image texture features in a compact and intuitive way. Our approach, called Dominant Texture Descriptor (DTD), is inspired by the dominant color descriptor. It is based on clustering the local texture features and identifying the dominant components and their spatial distribution. We also present an enhanced version of the DTD (eDTD) that encodes the spatial distribution of the pixels within each dominant component. We illustrate this approach for the case of two well-known descriptors, namely, the MPEG-7 Edge Histogram, and Ga- bor texture. The performance of the proposed texture feature representation is illustrated by using it to classify a collection of 900 color images. Experimental results are compared with those obtained using the traditional approaches. We show that our representation is more compact, interpretable, and could improve classification results by 10%-20%, especially for images with non-homogeneous texture.

We propose a content-based image retrieval prototype that combines the advantages of relevance feedback and image database categorization. Our approach is based on an algorithm that performs clustering and feature weighting simultaneously and can incorporate partial supervision information. This information, extracted from the userpsilas feedback through visual exploration and interaction, is used to refine the clusterspsila distributions and their feature relevance weights in the vicinity of the query image. The cluster dependent feature weights are used in the retrieval phase to adapt the similarity to the different categories. The feedback information is encoded as a set of constraints on which instances should or should not reside in the same cluster. Thus, it does not depend on the query image explicitly. Consequently, partial supervision information from different query sessions could be saved, accumulated, and used to continuously refine the image categories and their feature weights.

We propose a modified discrete HMM that includes a feature weighting discrimination component. We assume that the feature space is partitioned into subspaces and that the relevance weights of the different subspaces depends on the symbols and the states. In particular, we associate a partial probability with each symbol in each subspace. The overall observation state probability is then computed as an aggregation of the partial probabilities and their relevance weights. We consider two aggregation models. The first one is based on a linear combination, while the second one is based on a geometric combination. For both models, we reformulate the Baum-Welch learning algorithm and derive the update equations for the relevance weights and the partial state probabilities. The proposed approach is validated using synthetic and real data sets. The results are shown to outperform the baseline HMM.

We propose new continuous hidden Markov model (CHMM) structure that integrates feature weighting component. We assume that each feature vector could include different subsets of features that come from different sources of information or different feature extractors. We modify the probability density function that characterizes the standard CHMM to include state and component dependent feature relevance weights. To learn the optimal feature weights from the training data, we modify the maximum likelihood based Baum-Welch algorithm and we derive the necessary conditions. The proposed approach is validated using synthetic and real data sets. The results are shown to outperform the standard CHMM.

We present a novel method for fusing the results of multiple landmine detection algorithms that use different types of features, different classification methods, and different sensors. The proposed fusion method, called context-dependent multi-sensor fusion (CDMSF) is motivated by the fact that the relative performance of different detectors can vary significantly depending on the sensor, mine type, geographical site, soil and weather conditions, and burial depth. The training part of CDMSF has two components: context extraction and algorithm fusion. In context extraction, the features used by the different algorithms are combined and used to partition the feature space into groups of similar signatures, or contexts. The algorithm fusion component assigns an aggregation weight to each detector in each context based on its relative performance within the context. Results on ground penetrating radar (GPR) and wideband electromagnetic induction (WEMI) data collections show that the proposed method can identify meaningful and coherent clusters and that different expert algorithms can be identified for the different contexts. Our initial experiments have also indicated that the context-dependent fusion outperforms all individual detectors.

We introduce a new semi-supervised approach for Clustering and Aggregating Relational Data (SS-CARD). We assume that data is available in a relational form, where we only have information about the degrees to which pairs of objects in the data are related. Moreover, we assume that the relational information is represented by multiple dissimilarity matrices. These matrices could have been generated using different sensors, features, or mappings. The SS-CARD uses partial supervision information that consists of a small set of must-link and cannot-link constraints. The performance of the proposed algorithm is illustrated by using it to categorize a collection of 500 color images. The results are compared with those obtained by 3 other relational clustering methods.

In this paper, we introduce a semisupervised approach for clustering and aggregating relational data (SS-CARD). We assume that data is available in a relational form, where information only about the degrees to which pairs of objects in the dataset are related is available. Moreover, we assume that the relational information is represented bymultipledissimilarity matrices. These matrices could have been generated using different features, different mappings, or even different sensors. SS-CARD is designed to aggregate pairwise distances from multiple relational matrices, partition the data into clusters, and learn a relevance weight for each matrix in each cluster simultaneously. These weights have two main advantages. First, they help in partitioning the data into more meaningful clusters. Second, they can be used as part of a more complex learning system to enhance its learning behavior. SS-CARD uses partial supervision information that consists of a small set of constraints on which instances (shouldlink) or (shouldnotlink) reside in the same cluster. This additional information can guide the algorithm in learning the optimal relevance weights and in generating a better partition. The performance of the proposed algorithm is illustrated by using it in two different applications. The first one consists of categorizing the discrete nominal-valued mushroom data. The second application consists of categorizing a collection of images where each image is represented by several continuous features. For both applications, we represent the pairwise image dissimilarities by multiple relational matrices extracted from different feature sets. The results are compared with those obtained by three traditional relational clustering methods. We show that the partial supervision information and the learned aggregation weights can improve the results significantly.

In this paper, we address the problem of transforming relational features into an Euclidian space so that standard classification methods that assume that data is in a vector form could be used. Our approach has three main steps. First, a relational matrix that represents the pair-wise dissimilarities between all objects is constructed. Second, a fuzzy relational clustering algorithm is used to partition the data into groups of similar objects. Third, the relational data features are mapped to a unit hyper-cube space where each object is represented by its membership vectors in all clusters. The proposed method is validated by comparing the performance of several classifiers with different feature sets on the original and the transformed spaces. We show that the transformed space conserves the discriminative information of the original features. We also show that, using the transformed space, a richer set of standard classifiers could be used.

We propose a Semi-Supervised Clustering and Attribute Discrimination (S-SCAD) algorithm that performs fuzzy clustering and coarse feature weighting simultaneously. The supervision information in S-SCAD consists of a small set of constraints on which instances should or should not reside in the same cluster. The feature set is divided into logical subsets of features, and a degree of relevance is dynamically assigned to each subset based on its partial degree of dissimilarity. These weights have two advantages. First, they help in partitioning the data set into more meaningful clusters. Second, they can be used as part of a more complex learning system to enhance its learning behavior. We show that the partial supervision can guide the algorithm in learning the prototype parameters and the feature relevance weights, and thus, improve the final partition. The performance of the proposed algorithm is illustrated by using it to categorize a collection of color images. We use four feature subsets that encode color, structure, and texture information. The results are compared to other similar algorithms.

We propose an unsupervised approach to learn associations between continuous-valued attributes from different modalities. These associations are used to construct a multi-modal thesaurus that could serve as a foundation for inter-modality translation, and for hybrid navigation and search algorithms. We focus on extracting associations between visual features and textual keywords. Visual features consist of low-level attributes extracted from image content such as color, texture, and shape. Textual features consist of keywords that provide a description of the images. We assume that a collection of training images is available and that each image is globally annotated by few keywords. The objective is to extract representative visual profiles that correspond to frequent homogeneous regions, and to associate them with keywords. These profiles would be used to build the a multi-modal thesaurus. The proposed approach was trained with a large collection of images, and the constructed thesaurus was used to label new images. Initial experiments indicate that we can achieve up to 71.9% relative improvement on captioning accuracy over the state-of-the-art.

We introduce a new approach for Clustering and Aggregating Relational Data (CARD). We assume that data is available in a relational form, where we only have information about the degrees to which pairs of objects in the data set are related. Moreover, we assume that the relational information is represented by multiple dissimilarity matrices. These matrices could have been generated using different sensors, features, or mappings. CARD is designed to aggregate pairwise distances from multiple relational matrices, partition the data into clusters, and learn a relevance weight for each matrix in each cluster simultaneously. We introduce two versions of CARD. The first one is completely unsupervised(U-CARD). The second version is semi-supervised(SS-CARD) and uses partial supervision information that consists of a small set of must-link and cannot-link constraints. The performance of the proposed algorithms is illustrated by using it to categorize a collection of 500 color images. We represent the pairwise image dissimilarities by six different relational matrices that encode color, texture, and structure information. The results are compared with those obtained by 3 other relational clustering methods.

A variety of algorithms for the detection of landmines and discrimination between landmines and clutter objects have been presented. We discuss four quite different approaches in using data collected by a vehicle-mounted ground-penetrating radar sensor to detect landmines and distinguish them from clutter objects. One uses edge features in a hidden Markov model; the second uses geometric features in a feed-forward order-weighted average network; the third employs spectral features as its basis; and the fourth clusters edge histograms. We present the results of a large-scale cross-validation evaluation that uses a diverse set of data collected over 41 807.57$\hbox{m}^{2}$of ground, including 1593 mine encounters. Finally, we discuss the results of that ranking and what one can conclude concerning the performance of these four algorithms in various settings.

We propose an unsupervised approach to segment color images and annotate its regions. The annotation process uses a multi-modal thesaurus that is built from a large collection of training images by learning associations between low-level visual features and keywords. Association rules are learned through fuzzy clustering and unsupervised feature selection. We assume that a collection of images is available and that each image is globally annotated. The objective is to extract representative visual profiles that correspond to frequent homogeneous regions, and to associate them with keywords. Our approach has three main steps. First, each image is coarsely segmented into regions, and visual features are extracted from each region. Second, the regions are categorized using a fuzzy algorithm that performs clustering and feature weighting simultaneously. As a result, we obtain clusters of regions that share subsets of relevant features. Representatives from each cluster and their relevant visual and textual features would be used to build a thesaurus. Third, fuzzy membership functions are used to label new regions based on their proximity to the thesaurus entries. The proposed approach is trained with a collection of 2,695 images and tested with several different images.

This paper describes an algorithm for land mine detection using sensor data generated by a ground penetrating radar (GPR) system. The GPR produces a 3-D array of intensity values, representing a volume below the surface of the ground. First, a computationally inexpensive pre-screening algorithm is used to focus attention and identify regions with subsurface anomalies. The identified regions of interest are then processed by a feature extraction algorithm to capture their salient features. We use translation invariant features that are based on the local edge distribution of the 3-D GPR signatures. Finally, a fuzzy K-nearest neighbor rule is used to assign a confidence value to distinguish true detections from false alarms. The proposed algorithm is applied to data acquired from three outdoor test sites at different geographic locations.

We propose an unsupervised approach to segment color images and annotate its regions. The annotation process uses a multi-modal thesaurus that is built from a large collection of training images by learning associations between low-level visual features and keywords. We assume that a collection of images is available and that each image is globally annotated. The objective is to extract representative visual profiles that correspond to frequent homogeneous regions, and to associate them with keywords. These labeled profiles would be used to build a multi-modal thesaurus that could serve as a foundation for region based annotation. Our approach has two main steps. First, each image is coarsely segmented into regions, and visual features are extracted from each region. Second, the regions are categorized using a novel algorithm that performs clustering and feature weighting simultaneously. As a result, we obtain clusters of regions that share subsets of relevant features. Representatives from each cluster and their relevant visual and textual features would be used to build a thesaurus. The proposed approach is trained with a collection of 2695 images and tested with several different images.

We propose a new data-driven transformation that facilitates many data mining, interpretation, and analysis tasks. Our approach, called MembershipMap, strives to granulate and extract the underlying subconcepts of each raw attribute. The orthogonal union of these subconcepts are then used to define a new membership space. The subconcept soft labels of each point in the original space determine the position of that point in the new space. Since subconcept labels are prone to uncertainty inherent in the original data and in the initial extraction process, a combination of labeling schemes that are based on different measures of uncertainty will be presented. In particular, we introduce the CrispMap, the FuzzyMap, and the PossibilisticMap. We outline the advantages and disadvantages of each mapping scheme, and we show that the three transformed spaces are complementary. We also show that in addition to improving the performance of clustering by taking advantage of the richer information content, the MembershipMap can be used as a flexible preprocessing tool to support such tasks as: sampling, data cleaning, and outlier detection

We propose using a data preprocessing technique to automate the selection and labeling of points of interest from unlabeled data for the purpose of semisupervised clustering. The preprocessing approach, called MembershipMap, strives to extract the underlying sub-concepts of each attribute, and uses the orthogonal union of these sub-concepts to define a new, semantically richer, space. Using the MembershipMap, points of interest could be identified and labeled in a completely unsupervised way. We show that these points of interest could be used to convert an unsupervised learning problem into a simpler semi-supervised problem. Thus, benefiting from the advantage of semi-supervised learning without the need to manually identify and label a subset of control points. The proposed approach is applied to the problem of image segmentation

We propose a new data transformation approach that facilitates many data mining, interpretation, and analysis tasks. Our approach, called the membershipmap, strives to extract the underlying structure or sub-concepts of each raw attribute automatically, and uses the orthogonal union of these sub-concepts to define a new, semantically richer, space. The sub-concept labels of each point in the original space determine the position of that point in the transformed space. Since sub-concept labels are prone to uncertainty inherent in the original data and in the initial extraction process, a combination of labeling schemes that are based on different measures of uncertainty is presented. In particular, we introduce the crispmap, the fuzzymap, and the possibilisticmap. We outline the advantages and disadvantages of each mapping scheme, and we show that the three transformed spaces are complementary. The proposed transformation is illustrated with several data sets, and we show that it can be used as a flexible pre-processing tool to support such tasks as: sampling, data cleaning, and outlier detection.

This paper introduces a system for landmine detection using the sensor data generated by a ground penetrating radar (GPR). The GPR produces a three-dimensional array of intensity values, representing a volume below the surface of the ground. First, a constant false alarm rate (CFAR) detector is used to focus the attention and identify the candidates that resemble mines. Next, we apply a feature extraction algorithm based on projecting the data onto the dominant eigenvectors in the training data. The training signatures are then clustered to identify a few representatives, and a fuzzy k-nearest neighbor rule is used to distinguish true detections from false alarms.

Clustering is considered as one of the most important tools to organize and analyze large multimedia databases. Most existing clustering techniques assume that the clusters have well-defined shapes (spherical or ellipsoidal). Thus, they are not suitable for image database categorization where images are usually mapped to high-dimensional feature vectors, and it is hard to even guess the shape of the clusters in the feature space. In this paper, we assume that the high dimensional object signature can be modeled by a fuzzy set and we introduce an algorithm to cluster these sets. First, we define a measure to assess the dissimilarity between two fuzzy sets. Then, we integrate this measure into our synchronization-based clustering approach. The resulting algorithm, called SyMP/sub FD/ is robust to noise and outliers, determines the number of clusters in an unsupervised manner, and identifies clusters of arbitrary shapes. The robustness of SyMP/sub FD/ is an intrinsic property of the synchronization mechanism. To identify clusters of various shapes, SyMP/sub FD/ models each cluster by an ensemble of fuzzy sets. Clusters with simple shapes would be modeled by few sets while clusters with more complex shapes would require a larger number of sets. The performance of the proposed algorithm is illustrated by using it to categorize a collection of images, where each image is described by a fuzzy set representing its color distribution.

In this paper, we propose an algorithm that performs fuzzy clustering and feature weighting simultaneously and in an unsupervised manner. The feature set is divided into logical subsets of features, and a degree of relevance is dynamically assigned to each subset based on its partial degree of dissimilarity. The proposed algorithm is computationally and implementationally simple, and learns a different set of feature weights for each cluster. The cluster dependent feature weights have two advantages. First, they help in partitioning the data set into more meaningful clusters. Second, they can be used as part of a more complex learning system to enhance its learning behavior. The performance of the proposed algorithm is illustrated by using it to segment color images.

This paper introduces a new system for real-time land mine detection using sensor data generated by a Ground Penetrating Radar (GPR). The GPR produces a three-dimensional array of intensity values, representing a volume below the surface of the ground. Features are computed from this array and two types of membership degrees are assigned to each location. A fuzzy membership value provides a degree of belongingness of a given observation in the classes of mines, false alarms, and background, while a possibilistic membership value provides a degree of typicality. Both membership degrees are combined using simple rules to assign a confidence value. The parameters of the membership functions are obtained by clustering the training data and using the statistics of each partition. Our preliminary results show that the proposed approach is simple, efficient, and yet, yields results comparable to more complex detection systems.

We propose an approach to clustering text documents based on a coupled process of clustering and cluster-dependent keyword weighting. The proposed approach is based on the the fuzzy c-means clustering algorithm. Hence it is computationally and implementationally simple. Moreover, it learns a different set of keyword weights for each cluster. This means that, as a by-product of the clustering process, each document cluster will be characterized by a possibly different set of keywords. The cluster dependent keyword weights help in partitioning the document collection into more meaningful categories. They can also be used to automatically generate a brief summary of each cluster in terms of not only the attribute values, but also their relevance. For the case of text data, this approach can be used to automatically annotate the documents. We illustrate the performance of the proposed algorithm by using it to cluster a real collection of text documents.

We propose a new clustering algorithm, called SyMP, which is based on synchronization of pulse-coupled oscillators. SyMP represents each data point by an Integrate-and-Fire oscillator and uses the relative similarity between the points to model the interaction between the oscillators. SyMP is robust to noise and outliers, determines the number of clusters in an unsupervised manner, identifies clusters of arbitrary shapes, and can handle very large data sets. The robustness of SyMP is an intrinsic property of the synchronization mechanism. To determine the optimum number of clusters, SyMP uses a dynamic resolution parameter. To identify clusters of various shapes, SyMP models each cluster by multiple Gaussian components. The number of components is automatically determined using a dynamic intra-cluster resolution parameter. Clusters with simple shapes would be modeled by few components while clusters with more complex shapes would require a larger number of components. The scalable version of SyMP uses an efficient incremental approach that requires a simple pass through the data set. The proposed clustering approach is empirically evaluated with several synthetic and real data sets, and its performance is compared with CURE.

Traditional hand-written word recognition approaches use character recognition algorithms in the first stage that return confidences in each character under consideration for each segment in the hand-written word. These confidences are used in the second stage to compute a match score between the segmented hand-written word and a given character string. We explore an alternative approach where we view the first stage as a membership assignment process, rather than a character recognition process. The memberships are assigned based on the notion of typicality, and are not relative as in the case of a probabilistic framework. To generate the membership values, we use a robust clustering algorithm that can determine the number of prototypes required to model each character class in a robust and parsimonious manner. Each prototype represents a subclass of the character class. Our experimental results show that when used in conjunction with the proposed approach to word recognition, the memberships generated in this manner produce word recognition results that compare favorably with those of other methods.

This paper introduces a new scalable approach to clustering based on the synchronization of pulse-coupled oscillators. Each data point is represented by an integrate-and-fire oscillator and the interaction between oscillators is defined according to the relative similarity between the points. The set of oscillators self-organizes into stable phase-locked subgroups. Our approach proceeds by loading only a subset of the data and allowing it to self-organize. Groups of synchronized oscillators are then summarized and purged from memory. We show that our method is robust, scales linearly and can determine the number of clusters. The proposed approach is empirically evaluated with several synthetic data sets and is used to segment large color images.

We introduce an efficient synchronization model that organizes a population of integrate-and-fire oscillators into stable and structured groups. Each oscillator fires synchronously with all the others within its group, but the groups themselves fire with a constant phase difference. The structure of the synchronized groups depends on the choice of the coupling function. We show that by defining the interaction between oscillators according to the relative distance between them, our model can be used as a general clustering algorithm. Unlike existing models, our model incorporates techniques from relational and prototype-based clustering methods and results in a clustering algorithm that is simple, efficient, robust, unbiased by the size of the clusters, and that can find an arbitrary number of clusters. In addition to helping the model self-organize into stable groups, the synergy between clustering and synchronization reduces the computational complexity significantly. The resulting clustering algorithm has several advantages over conventional clustering techniques. In particular, it can generate a nested sequence of partitions and it can determine the optimum number of clusters in an efficient manner. Moreover, since our approach does not involve optimizing an objective function, it is not sensitive to initialization and it can incorporate nonmetric similarity measures. We illustrate the performance of our algorithms with several synthetic and real data sets.

We propose a new algorithm that performs clustering and feature weighting simultaneously and in an unsupervised manner. The proposed algorithm is computationally and implementationally simple, and learns a different set of feature weights for each cluster. The cluster dependent feature weights have two advantages. First, they help in partitioning the data set into more meaningful clusters. Second, they can be used as part of a more complex learning system to enhance its learning behavior. The performance of the proposed algorithm is illustrated by using it to segment real color images.

This paper addresses three major issues associated with conventional partitional clustering, namely, sensitivity to initialization, difficulty in determining the number of clusters, and sensitivity to noise and outliers. The proposed robust competitive agglomeration (RCA) algorithm starts with a large number of clusters to reduce the sensitivity to initialization, and determines the actual number of clusters by a process of competitive agglomeration. Noise immunity is achieved by incorporating concepts from robust statistics into the algorithm. RCA assigns two different sets of weights for each data point: the first set of constrained weights represents degrees of sharing, and is used to create a competitive environment and to generate a fuzzy partition of the data set. The second set corresponds to robust weights, and is used to obtain robust estimates of the cluster prototypes. By choosing an appropriate distance measure in the objective function, RCA can be used to find an unknown number of clusters of various shapes in noisy data sets, as well as to fit an unknown number of parametric models simultaneously. Several examples, such as clustering/mixture decomposition, line/plane fitting, segmentation of range images, and estimation of motion parameters of multiple objects, are shown.

We present a new clustering algorithm that addresses two major issues associated with conventional partitional clustering: the difficulty in determining the number of clusters, and the sensitivity to noise and outliers. The proposed algorithm determines the number of clusters by a process of competitive agglomeration. Noise immunity is achieved by integrating concepts from robust statistics into the algorithm. The proposed approach can incorporate different distance measures in the objective function to find an unknown number of clusters of various types including lines, planes and surfaces.

In this paper, we introduce a shell-clustering algorithm for ellipsoidal clusters based on the so-called "radial distance" which can be easily extended to superquadric clusters. We compare our algorithm with other algorithms in the literature that are based on the algebraic distance, the approximate distance, the normalized radial distance, and the exact distance. We evaluate the performance of each algorithm on two-dimensional data sets containing "scattered" ellipses, partial ellipses, outliers, and ellipses of disparate sizes, and summarize the relative strengths and weaknesses of each algorithm.

Traditionally, prototype-based fuzzy clustering algorithms such as the Fuzzy C Means (FCM) algorithm have been used to find "compact" or "filled" clusters. Recently, there have been attempts to generalize such algorithms to the case of hollow or "shell-like" clusters, i.e., clusters that lie in subspaces of feature space. The shell clustering approach provides a powerful means to solve the hitherto unsolved problem of simultaneously fitting multiple curves/surfaces to unsegmented, scattered and sparse data. In this paper, we present several fuzzy and possibilistic algorithms to detect linear and quadric shell clusters. We also introduce generalizations of these algorithms in which the prototypes represent sets of higher-order polynomial functions. The suggested algorithms provide a good trade-off between computational complexity and performance, since the objective function used in these algorithms is the sum of squared distances, and the clustering is sensitive to noise and outliers. We show that by using a possibilistic approach to clustering, one can make the proposed algorithms robust.>

Shell clustering algorithms are ideally suited for computer vision tasks such as boundary detection and surface approximation, particularly when the boundaries have jagged or scattered edges and when the range data is sparse. This is because shell clustering is insensitive to local aberrations, it can be performed directly in image space, and unlike traditional approaches it does assume dense data and does not use additional features such as curvatures and surface normals. The shell clustering algorithms introduced in Part I of this paper assume that the number of clusters is known, however, which is not the case in many boundary detection and surface approximation applications. This problem can be overcome by considering cluster validity. We introduce a validity measure called surface density which is explicitly meant for the type of applications considered in this paper, we show through theoretical derivations that surface density is relatively invariant to size and partiality (incompleteness) of the clusters. We describe unsupervised clustering algorithms that use the surface density measure and other measures to determine the optimum number of shell clusters automatically, and illustrate the application of the proposed algorithms to boundary detection in the case of intensity images and to surface approximation in the case of range images.>

The fuzzy c spherical shells (FCSS) algorithm is specially designed to search for clusters that can be described by circular arcs or, generally, by shells of hyperspheres. A new approach to the FCSS algorithm is presented. This algorithm is computationally and implementationally simpler than other clustering algorithms that have been suggested for this purpose. An unsupervised algorithm which automatically finds the optimum number of clusters is not known. It uses a cluster validity measure to identify good clusters, merges all compatible clusters, and eliminates spurious clusters to achieve the final results. Experimental results on several data sets are presented.