While deep learning approach has been prevalent for generating image features, conventional handcrafted salience features still have the strength of providing explicit domain knowledge and reflecting intuitive visual understanding. However, the existing usages of handcrafted features in deep network are lack in addressing the issues of parameter quality. In this research, we propose a novel pooling method that enriches the deep features by utilizing the injected salience shape features - Generic Edge Tokens and Curve Partitioning Points, to adjust the outputs of pooling layer. The model trained under the guidance of domain prior knowledge is able to produce deep representation embracing merits from both handcrafted features and deep network. The experimental results show its improved performance with reduced learning curve. The proposed novel pooling method is generic, ie. open to other handcrafted features and different deep network architectures.

Learning from imbalanced datasets is challenging due to the existence of imbalanced class distribution and other problems such as class overlapping, high dimensionality of the dataset and small sample size. This paper focuses on handling the class imbalanced problem with sparsity. A hybrid sampling method of “Under-sampling + Over-sampling” is proposed. Under-sampling selects informative instances and features from the original dataset, whereas Over-sampling balances the under-sampled dataset. Considering that safe double screening can quickly identify and remove non-informative instances and features from the dataset with sparse structure, it is adopted as Under-sampling in the hybrid sampling method. Under-sampling based on safe double screening can obtain all of the boundary instances and the informative features for classification by utilizing sparse structure of datasets. Experimental results show that safe double screening can reduce the number of instances and features and capture the true imbalance ratio of a dataset. The classifiers built on the sampled datasets by using the hybrid sampling method obtain better classification performance, especially for the data in the minority class. This hybrid sampling method is suitable for constructing a classifier based on decision boundary on sparse imbalanced dataset in which the number of features is large.

In this research, we propose an object localization method to boost the performance of current object recognition techniques. This method utilizes the image edge information as a clue to determine the location of the objects. The Generic Edge Tokens (GETs) of the image are extracted based on the perceptual organization elements of human vision. These edge tokens are parsed according to the Best First Search algorithm to fine-tune the location of objects, where the objective function is the detection score returned by the Deep Convolutional Neural Network. We have evaluated our method on top of the RCNN object recognition method. The results on Pascal VOC 2007 and 2012 show improved object localization performance. We also present several cases where the proposed method works significantly more precisely than RCNN.

Image classification is a general visual analysis task based on the image content coded by its representation. In this research, we proposed an image representation method that is based on the perceptual shape features and their spatial distributions. A natural language processing concept, N-gram, is adopted to generate a set of perceptual shape visual words for encoding image features. By combining hierarchical visual words and spatial pyramid, Spatio-Shape Pyramid representation is constructed to reduce the semantic gaps. Experimental results show that the proposed method outperforms other state-of-the-art methods.

Despite the outperformance of Support Vector Machine (SVM) on many practical classification problems, the algorithm is not directly applicable to multi-dimensional trajectories having different lengths. In this paper, a new class of SVM that is applicable to trajectory classification, such as action recognition, is developed by incorporating two efficient time-series distances measures into the kernel function. Dynamic Time Warping and Longest Common Subsequence distance measures along with their derivatives are employed as the SVM kernel. In addition, the pairwise proximity learning strategy is utilized in order to make use of non-positive semi-definite kernels in the SVM formulation. The proposed method is employed for a challenging classification problem: action recognition by depth cameras using only skeleton data; and evaluated on three benchmark action datasets. Experimental results demonstrate the outperformance of our methodology compared to the state-of-the-art on the considered datasets.

Multiple classifier systems, also known as classifier ensembles, have received great attention in recent years because of their improved classification accuracy in different applications. In this paper, we propose a new general approach to ensemble classification, named generic subclass ensemble, in which each base classifier is trained with data belonging to a subset of classes, and thus discriminates among a subset of target categories. The ensemble classifiers are then fused using a combination rule. The proposed approach differs from existing methods that manipulate the target attribute, since in our approach individual classification problems are not restricted to two-class problems. We perform a series of experiments to evaluate the efficiency of the generic subclass approach on a set of benchmark datasets. Experimental results with multilayer perceptrons show that the proposed approach presents a viable alternative to the most commonly used ensemble classification approaches.

The performance of different action-recognition methods using skeleton joint locations have been recently studied by several computer vision researchers. However, the potential improvement in classification through classifier fusion by ensemble-based methods has remained unattended. In this work, we evaluate the performance of an ensemble of five action learning techniques, each performing the recognition task from a different perspective. The underlying rationale of the fusion approach is that different learners employ varying structures of input descriptors/features to be trained. These varying structures cannot be attached and used by a single learner. In addition, combining the outputs of several learners can reduce the risk of an unfortunate selection of a poorly performing learner. This leads to having a more robust and general-applicable framework. Also, we propose two simple, yet effective, action description techniques. In order to improve the recognition performance, a powerful combination strategy is utilized based on the Dempster-Shafer theory, which can effectively make use of diversity of base learners trained on different sources of information. The recognition results of the individual classifiers are compared with those obtained from fusing the classifiers' output, showing advanced performance of the proposed methodology.

Multiple classifier systems, also known as classifier ensembles, have received great attention in recent years because of the improved classification accuracy in different applications. A large variety of ensemble methods have been proposed in order to exploit strengths of individual classifiers. In this paper, we present a unifying framework for multiple classifier systems, which unites most classification methods by an ensemble of classifiers. Specifically, we link two research lines in machine learning: multiclass classification based on the class binarization techniques and the strategies of ensemble classification. With the proposed framework, the various ensemble classification strategies will be broadly categorized into four main approaches. Then, we provide a brief survey of ensemble methods based on these main approaches as well as principle techniques proposed to combine them.

Among the proposed methods to deal with multi-class classification problems, the Error-Correcting Output Codes (ECOC) represents a powerful framework. The key factor in designing any ECOC matrix is the independency of the binary classifiers, without which the ECOC method would be ineffective. This paper proposes an efficient new approach to the ECOC framework in order to improve independency among classifiers. The underlying rationale for our work is that we design three-dimensional codematrix, where the third dimension is the feature space of the problem domain. Using rough set-based feature selection, a new algorithm, named "Rough Set Subspace ECOC (RSS-ECOC)" is proposed. We introduce the Quick Multiple Reduct algorithm in order to generate a set of reducts for a binary problem, where each reduct is used to train a dichotomizer. In addition to creating more independent classifiers, ECOC matrices with longer codes can be built. The numerical experiments in this study compare the classification accuracy of the proposed RSS-ECOC with classical ECOC, one-versus-one, and one-versus-all methods on 24 UCI datasets. The results show that the proposed technique increases the classification accuracy in comparison with the state of the art coding methods.

Generic shape feature extraction is a challenging task for image and video content analysis. We present a non-parametric statistics based method for extracting generic shape tokens based on a Perceptual Curve Partition and Grouping (PCPG) model. In this PCPG model, each curve is made up of Generic Edge Tokens (GET) connected at Curve Partitioning Points (CPP). The types of GET and CPP provide a set of basic shape descriptors for semantic vocabulary. The new implementation of the PCPG is based on: 1) An arctangent space is employed to signify the evidence of CPPs at pixel-level. 2) The pixels' sequential order is taken as heuristic to establish a bin order preserving arctangent histogram for locating CPPs by examining the continuity of generic feature criteria statistically. 3) A new CPP detection scheme is capable of detecting CPPs and classifying GETs on the fly. Experiments are presented for performance demonstration.

In this paper, we present a new technique, called stream projected ouliter detector (SPOT), to deal with outlier detection problem in high-dimensional data streams. SPOT is unique in a number of aspects. First, SPOT employs a novel window-based time model and decaying cell summaries to capture statistics from the data stream. Second, sparse subspace template (SST), a set of top sparse subspaces obtained by unsupervised and/or supervised learning processes, is constructed in SPOT to detect projected outliers effectively. Multi-Objective genetic algorithm (MOGA) is employed as an effective search method in unsupervised learning for finding outlying subspaces from training data. Finally, SST is able to carry out online self- evolution to cope with dynamics of data streams. This paper provides details on the motivation and technical challenges of detecting outliers from high-dimensional data streams, present an overview of SPOT, and give the plans for system demonstration of SPOT.

In this paper, we study the problem of anomaly detection in high-dimensional network streams. We have developed a new technique, called Stream Projected Outlier deTector (SPOT), to deal with the problem of anomaly detection from high-dimensional data streams. We conduct a case study of SPOT in this paper by deploying it on 1999 KDD Intrusion Detection application. Innovative approaches for training data generation, anomaly classification and false positive reduction are proposed in this paper as well. Experimental results demonstrate that SPOT is effective in detecting anomalies from network data streams and outperforms existing anomaly detection methods.

Detecting outlying subspaces is a relatively new research problem in outlier-ness analysis for high-dimensional data. An outlying subspace for a given data point p is the subspace in which p is an outlier. Outlying subspace detection can facilitate a better characterization process for the detected outliers. It can also enable outlier mining for highdimensional data to be performed more accurately and efficiently. In this paper, we proposed a new method using genetic algorithm paradigm for searching outlying subspaces efficiently. We developed a technique for efficiently computing the lower and upper bounds of the distance between a given point and its kth nearest neighbor in each possible subspace. These bounds are used to speed up the fitness evaluation of the designed genetic algorithm for outlying subspace detection. We also proposed a random sampling technique to further reduce the computation of the genetic algorithm. The optimal number of sampling data is specified to ensure the accuracy of the result. We show that the proposed method is efficient and effective in handling outlying subspace detection problem by a set of experiments conducted on both synthetic and real-life datasets.

The problem of gene specific co-regulation discovery is that, for a particular gene of interest, identify its closely coregulated genes and the associated subsets of experimental conditions in which such co-regulations occur. The coregulations are local in the sense that they occur in some subsets of full experimental conditions. In this paper, we propose an innovative method for finding gene specific coregulations using genetic algorithm (GA). Two novel ad hoc GAs, the single-stage and two-stage progressive GA, are proposed. They are called progressive because the initial population for the GA in a window position inherits the top-ranked individuals obtained in the preceding window position, enabling them to achieve better accuracy than the nonprogressive algorithm. Experimental results with real-life gene expression data demonstrate the efficiency and effectiveness of our technique in discovering gene specific coregulations

This paper presents a method based on perceptual organization for tracking object motion within image sequences (or video). To achieve the characteristics of real-time, efficiency and robustness, a perceptual edge tracker was developed to extract generic edge tokens (GETS) on the fly. The GETS are defined qualitatively according to their descriptive properties. Object motion is tracked within the image sequence by computing the difference between GET streams in consecutive frames. Multiple moving objects are grouped by forming clusters of motion GET. Experimental results are provided.

This paper presents a generic approach of visual knowledge representation for man-made object recognition based on perceptual organization. In the approach, a generic model scheme is developed for coding both perceptual structures of objects and generic viewing situations of the objects in a qualitative manner. A set of generic edge features are defined as perceptual tokens for constructing structures of generic models. Faces are used as basic perceptual entities of a generic model. A synthetic view space is developed for coding generic views of the models and grouping perceptual structures based on the principles of shape constancy and face-pose consistency.

The visual appearance of an object in space is an image configuration projected from a subset of connected faces of the object. It is believed that face perception and face integration play a key role in object recognition in human vision. This paper presents a novel approach for calculating viewpoint consistency for three-dimensional (3D) object recognition, which utilizes the perceptual models of face grouping and face integration. In the approach, faces are used as perceptual entities in accordance with the visual perception of shape constancy and face-pose consistency. To accommodate the perceptual knowledge of face visibility of objects, a synthetic view space (SVS) is developed. SVS is an abstractive perceptual space which partitions and synthesizes the conventional metric view sphere into a synthetic view box in which only a very limited set of synthetic views (s-views) need to be considered in estimating face-pose consistency. The s-views are structurally organized in a network, the view-connectivity net (VCN), which describes all the possible connections and constraints of the s-views in SVS. VCN provides a meaningful mechanism in pruning the search space of SVS during estimating face-pose consistency. The method has been successfully used for recognizing a class of industrial parts.