Researchers have been involved for decades in search of an efficient skin detection method. However, current methods have not overcome the significant challenges of skin detection, such as variation of illumination, various skin tones of different ethnic groups, and many others. This research proposed a clustering and region-growing-based skin detection method to overcome these limitations. Together with significant insight, these methods result in a more effective algorithm. The insight concerns the capability to dynamically define the number of clusters in a collection of pixels organized as images. In Clustering for most problem domains, the number of clusters is fixed prior and does not perform effectively over a wide variety of data contents. Therefore, this research paper proposed a skin detection method that validated the above findings. The proposed method assigns the number of clusters based on image properties and ultimately allows freedom from manual thresholds or other manual operations. The dynamic determination of clustering outcomes allows for greater automation of skin detection when dealing with uncertain real-world conditions.

Action recognition is one of the most important components for video analysis. In addition to objects and atomic actions, temporal relationships are important characteristics for many actions and are not fully exploited in many approaches. We model the temporal structures of midlevel actions (referred to as components) based on dense trajectory components, obtained by clustering individual trajectories. The trajectory components are a higher level and a more stable representation than raw individual trajectories. Based on the temporal ordering of trajectory components, we describe the temporal structure using Allen's temporal relationships in a discriminative manner and combine it with a generative model using bag of components. The main idea behind the model is to extract midlevel features from domain-independent dense trajectories and classify the actions by exploring the temporal structure among these midlevel features based on a set of relationships. We evaluate the proposed approach on public data sets and compare it with a bag-of-words–based approach and state-of-the-art application of the Markov logic network for action recognition. The results demonstrate that the proposed approach produces better recognition accuracy.

In this paper we present an automatic vehicle ingress/egress counting method by clustering dense trajectories extracted from monitoring videos. Dense trajectories are extracted based on dense optical flow when passengers cross the door of the vehicle, and then clustered into different groups according to their descriptors with each legitimate group as a passenger. The contribution of the proposed method is twofold. First, we put forward an online passenger counting framework which is based on feature-points tracking and can be easily deployed to different scenarios. The method works even in low illumination conditions as demonstrated in experiments. Second, vehicle running information was combined to improve the accuracy of passenger counting. The transit vehicle settings are unconstrained and complex due to variations from illumination, movement and uncontrolled passenger behaviors. We tackle this by incorporating different modalities besides videos such as the status of the vehicle (e.g., in motion or not). The experimental results on multiple real bus videos show that the proposed system can count passengers with average accuracy of 94.9% at an average frame rate of 38 fps.

Since wireless capsule endoscopy (WCE) is a novel technology for recording the videos of the digestive tract of a patient, the problem of segmenting the WCE video of the digestive tract into subvideos corresponding to the entrance, stomach, small intestine, and large intestine regions is not well addressed in the literature. A selected few papers addressing this problem follow supervised leaning approaches that presume availability of a large database of correctly labeled training samples. Considering the difficulties in procuring sizable WCE training data sets needed for achieving high classification accuracy, we introduce in this paper an unsupervised learning approach that employs Scale Invariant Feature Transform (SIFT) for extraction of local image features and the probabilistic latent semantic analysis (pLSA) model used in the linguistic content analysis for data clustering. Results of experimentation indicate that this method compares well in classification accuracy with the state-of-the-art supervised classification approaches to WCE video segmentation.

A novel method for three-dimensional (3D) residential building reconstruction in urban areas using LiDAR (light detection and ranging) data is proposed. The main contribution of this work is the automatic segmentation of roof points and roof type recognition and reconstruction based on sparse LiDAR data. Using minimum bounding rectangle (MBR) method and model-based reconstruction, we are able to automatically identify individual buildings from cluttered residential areas and re-create building models with improved accuracy in a reasonably short time. We applied our method to urban sites in the city of New Orleans and demonstrated that the method identified building measurements successfully from LiDAR data and rebuilt 3D models effectively. Our experiments show that even in the presence of noise we can successfully reconstruct small buildings given relatively sparse LiDAR samples with help from template databases.

Tagline detection and indexing are challenging tasks due to complicated anatomical properties and imaging noise. In this paper, we will address the following two important issues in tagline detection: 1) an automatic method independent from imaging approaches with improved robustness and accuracy and 2) tagline indexing that matches taglines in task and reference images for postprocessing. Our method consists of two steps: First, a wavelet decomposition is performed on a tagged magnetic resonance (tMR) image. Subband correlation is used to dampen anatomical boundaries but enhance taglines. A tagline map is created by segmenting a reconstructed image using pseudowavelet reconstruction. Next, tagline pixels are grouped into clusters and isolated small line segments are eliminated. A snake method is then used to index and recover broken taglines. Our method has been validated with 320 tMR tongue images. Measurement of tagline accuracy was performed by computing tag pixel displacement. Without assumptions on tagline models, it detects taglines automatically. Comparison studies were conducted against the harmonic phase method. Our experiments resulted in a p-value of 1E-6 with one-way ANOVA, which indicates a significant improvement in accuracy and robustness.

We present a preprocessing method for automatic break line detection. Our method is given a set of edges (break lines and non-break lines) we eliminate the non-break line edges leaving only those with higher probability of being break lines. Our method is based on fusing IR images and LiDAR cloud points. In the first step, we apply the Canny edge detection algorithm to the IR images (producing a superset of break lines). Then we project the LiDAR points onto a 2-D plane, ignoring the set of points that are greater than a selected threshold (different elevation thresholds have been selected), which allows the footprints of some elevated structures to appear clearly in the set of projected points. Those structures that appear in both LiDAR points and IR images are used as references for registration of LiDAR cloud points with the IR images. After registration, we eliminate all the edges that appear in flat areas, which is achieved by applying a 3D filter to the LiDAR points.

To generate a DTM, measurements from above-ground features such as buildings, vehicles, and vegetation have to be classified and removed, which is nontrivial. The above-ground features present great challenges in conjunction with varying slopes of the ground. In this paper, we present a method to remove above-ground LiDAR measurements and generate DTMs by using adaptive window size according to the local gradients. Iterative construction measurements are performed until difference between two iterations are minimum. In our experiments, we apply our method to the LiDAR data acquired from the downtown region of New Orleans. It was demonstrated that the adaptive window method can remove most of the above-ground points effectively.

Fusion of images in the presence of noise is a challenging problem. Conventional fusion methods focus on aggregating prominent image features, which usually result in noise enhancement. To address this problem, we developed a wavelet-based, noise-aware fusion method that distinguishes signal and noise coefficients on-the-fly and fuses them with weighted averaging and majority voting respectively. Our method retains coefficients that reconstruct salient features, whereas noise components are discarded. The performance is evaluated in terms of noise removal and feature retention. The comparisons with five state-of-the-art fusion methods and a combination with denoising method demonstrated that our method significantly outperformed the existing techniques with noisy inputs.

There has been increasing interest to design better probability estimation trees, or PETs, for ranking and probability estimation. Capable of generating class membership probabilities, PETs have been shown to be highly accurate and flexible for many difficult problems, such as cost-sensitive learning and matching skewed distributions. There are a large number of PET algorithms available, and about ten of them are well-known. This large number provides an advantage, but it also creates confusion in practice. One would ask "given a new dataset, which algorithm to choose and what performance to expect and not to expect? What are the reasons to explain either good or bad performance under different situations?" In this paper, we systematically, for the first time, answer these important questions by conducting a large-scale empirical comparison of five popular PETs by examining their AUC, MSE and error rate "learning curves" (instead of training-test split based cross-validation). Using the maximum AUC achieved by any of the evaluated probability estimation tree algorithms, we demonstrate that the preference of a probability estimation tree on different evaluation metrics can be accurately characterized by the "signal-noise separability" of the dataset, as well as some other observable statistics of the dataset explained further in the paper. Moreover, in order to understand their relative performance, many important and previously unrevealed properties of each PET's mechanism and heuristics are analyzed and evaluated. Importantly, a practical guide for choosing the most appropriate PET algorithm given a new data mining problem is provided.

In practice, learning from data is often hampered by the limited training examples. In this paper, as the size of training data varies, we empirically investigate several probability estimation tree algorithms over eighteen binary classification problems. Nine metrics are used to evaluate their performances. Our aggregated results show that ensemble trees consistently outperform single trees. Confusion factor trees(CFT) register poor calibration even as training size increases, which shows that CFTs are potentially biased if data sets have small noise. We also provide analysis on the observed performance of the tree algorithms.

Niching techniques diversify the population of evolutionary algorithms, encouraging heterogeneous convergence to multiple optima. The key to an effective diversification is identifying the similarity among individuals. With no prior knowledge of the fitness landscapes, it is usually determined by uninformative assumptions on the number of peaks. We propose a method to estimate the sharing distance and the corresponding population size. Using the probably approximately correct (PAC) learning theory and the e-cover concept, we derive the PAC neighbor distance of a local optimum. Within this neighborhood, uniform samples are drawn and we compute the subspace fitness distance correlation (FDC) coefficients. An algorithm is developed to estimate the granularity feature of the fitness landscapes. The sharing distance is determined from the granularity feature and furthermore, the population size is decided. Experiments demonstrate that by using the estimated population size and sharing distance an evolutionary algorithm (EA) correctly identifies multiple optima.

An approach to fuse multiple images based on fuzzy feature representation and evidential reasoning is proposed in this article. Fuzzy set and Dempster-Shafer theory provides a complete framework to describe information naturally and combine weak evidence from multiple sources. Such situations typically arise in the image fusion problems, where a 'real scene' image has to be estimated from incomplete and unreliable observations. By converting images from their spatial domain into the fuzzy evidential representations, decisions are made to aggregate evidence such that a fused image is generated. The proposed fusion approach is evaluated on a broad set of images and promising results are given.

The focus of this paper is the discovery of negative association rules. Such association rules are complementary to the sorts of association rules most often encountered in the literature and have the forms of X/spl rarr/ -Y or -X/spl rarr/Y. We present a rule discovery algorithm that finds a useful subset of valid negative rules. In generating negative rules, we employ a hierarchical graph-structured taxonomy of domain terms. A taxonomy containing classification information records the similarity between items. Given the taxonomy, sibling rules, duplicated from positive rules with a couple of items replaced, are derived together with their estimated confidence. Those sibling rules that bring big confidence deviation are considered candidate negative rules. Our study shows that negative association rules can be discovered efficiently from large database.

This paper presents a genetic programming approach for the synthesis of logic functions by means of multiplexers. The approach uses the 1-control line multiplexer as the only design unit. Any logic function (defined by a truth table) can be produced through the replication of this single unit. Our fitness function works in two stages: first, it finds feasible solutions, and then it concentrates on the minimization of the circuit, The proposed approach does not require any knowledge from the application domain.

Probably approximately correct learning, PAC-learning, is a framework for the study of learnability and learning machines. In this framework, learning is induced through a set of examples. The size of this set is such that with probability greater than 1-/spl delta/ the learning machine shows an approximately correct behavior with error no greater than /spl epsiv/. The authors use the PAC framework to derive the size of a GA population that with probability 1-/spl delta/ contains at least one individual /spl epsiv/-close to a target hypothesis or solution.

Use of a genetic algorithm assumes the existence of a figure of merit called fitness, for which there is a value for every candidate solution. The fitness must be measurable over the representation of the solution by means of a computable function. The fitness function is, in most cases, independent of the other factors, including the algorithm used. Often, the fitness is an estimation of the nearness to an ideal solution or the distance from a default solution. In image scene interpretation, the solution takes the form of a set of labels corresponding to the components of an image and its fitness is difficult to conceptualize in terms of distance from a default or nearness to an ideal. Here we describe a model in which a semantic net is used to capture the salient properties of an ideal labeling. Instantiating the nodes of the semantic net with the labels from a candidate solution (a chromosome) provides a basis for estimating a logical distance from a norm. This domain-independent model can be applied to a broad range of scene-based image analysis tasks. © 2000 John Wiley & Sons, Inc.

Niching or speciation is a particularly appropriate for multimodal function optimization. An irony in EC research is that genetic algorithms (GAs) are not touted primarily as function optimizers yet all reported niching research is in the GA context. Evolution strategies (ES) and major variants of evolutionary programming are better suited by design for global optimization. Borrowing methods that have been reported for niching in GAs, we have applied them to an EC algorithm that resembles ES in structure. We have found that the selection methods in ES, e.g., (/spl mu/+/spl lambda/), interact satisfactorily with niching strategies used in GAs. On the other hand, adapting selection methods such as SUS that minimize bias does not lead to favorable results. This is counter to expectations but can be reconciled with prevailing theories. We conclude with a conjecture concerning a lower bound on population size for multimodal optimization.

This paper presents an approach to the experimental verification of the quality of "model selection" delivered by statistical learning theory (SLT). We depart from a function whose analytical approximation properties by polynomials are well known and readily verifiables in our experimental environment. For different sample size sets, the model predicted by SLT is contrasted against the model derived from the mathematical properties of the function. We found great robustness in the predictive ability of SLT.

Trends in databases leading to complex objects present opportunities for representing imprecision and uncertainty that were difficult to integrate cohesively in simpler database models. In fact, one can begin at the conceptual level with a model that allows uncertainty assumptions and then transform those assumptions into a logical model having the necessary semantic foundations upon which to base a meaningful query language. Here we provide such a constructive approach beginning with the ExIFO model for expression of the conceptual design then show how the conceptual design is transformed into the logical design (for which we utilize the extended NF/sup 2/ logical database model). The steps are straightforward, unambiguous, and preserve the relevant information, including information concerning uncertainty.

There have been significant theoretical advances in fuzzy database technology, yet commercially its successes have been negligible. This article examines the current state of this technology and suggests directions for future efforts. A framework for the analysis of fuzzy database technology is proposed and extant models are examined with reference to this framework. Fuzzy databases are studied in relation to the requirements of the database community. It is argued that new generation applications and object-oriented databases hold the key to the future commercial acceptability of this technology. © 1996 John Wiley & Sons, Inc.

This paper fully develops a previous approach by George et al. (1993) to modeling uncertainty in class hierarchies. The model utilizes fuzzy logic to generalize equality to similarity which permitted impreciseness in data to be represented by uncertainty in classification. In this paper, the data model is formally defined and a nonredundancy preserving primitive operator, the merge, is described. It is proven that nonredundancy is always preserved in the model. An object algebra is proposed, and transformations that preserve query equality are discussed.

This paper describes a database model based on the original rough sets theory. Its rough relations permit the representation of a rough set of tuples not definable in terms of the elementary classes, except through use of lower and upper approximations. The rough relational database model also incorporates indiscernibility in the representation and in all the operators of the rough relational algebra. This indiscernibility is based strictly on equivalence classes which must be defined for every attribute domain. There are several obvious applications for which the rough relational database model can more accurately model an enterprise than does the standard relational model. These include systems involving ambiguous, imprecise, or uncertain data. Retrieval over mismatched domains caused by the merging of one or more applications can be facilitated by the use of indiscernibility, and naive system users can achieve greater recall with the rough relational database. In addition, applications inherently “rough” could be more easily implemented and maintained in the rough relational database.

Dataflow graphs are a generalized model of computation. Uninterpreted dataflow graphs with nondeterminism resolved via probabilities are shown to be isomorphic to a class of Petri nets known as free choice nets. Petri net analysis methods are readily available in the literature and this result makes those methods accessible to dataflow research. Nevertheless, combinatorial explosion can render Petri net analysis inoperative. Using a previously known technique for decomposing free choice nets into smaller components, it is demonstrated that, in principle, it is possible to determine aspects of the overall behavior from the particular behavior of components.

In this paper, a new model for parallel computations and parallel computer systems that is based on data flow principles is presented. Uninterpreted data flow graphs can be used to model computer systems including data driven and parallel processors. A data flow graph is defined to be a bipartite graph with actors and links as the two vertex classes. Actors can be considered similar to transitions in Petri nets, and links similar to places. The nondeterministic nature of uninterpreted data flow graphs necessitates the derivation of liveness conditions.

A fuzzy relational database is a medium capable of representing information that is inherently imprecise or is the aggregation of the subjective opinions of a number of individuals. Measuring the degree of precision or lack thereof is important for two reasons. First, if the measures themselves achieve extrema when conditions match what is intuitively recognized as maximum and minimum fuzziness, then confidence in the medium to faithfully represent the intervening range of precision is increased. Second, querying fuzzy information may result in ambiguous replies and measures of preciseness may fathom how well the query discriminated among the possible replies. Entropy measures based on the fuzzy and probabilistic attributes of databases are developed and applied to the two ends mentioned above.

Future ballistic missile defense systems will involve multiple computers combined in complex ways. Prototype and breadboard evaluation will be prohibitively expensive if more than one alternative is investigated. Nevertheless, confidence must be gained in the hardware before it is actually constructed and the software must be available much earlier in the development cycle. BMD Advanced Technology Center is investigating a comprehensive emulation facility to accomplish both these goals. This paper addresses the shortcomings of current technology, characterizes the problems in extending emulation to encompass these goals, and provides an initial basis for solution.

Let C = {C1, C2,..., Cm} be the set of combinations of n items taken p at a time and arranged in lexicographical order.Given an integer, i (1 ~ i __< m, n >__ p ~ 0), the algorithm derives C,.Previous algorithms [1,3, 6] have accomplished this by the generation of all vectors between an initial point in the combination space and the desired vector.The cited algorithms are computationally advantageous if all combinations or a sequential subset are required.However, the algorithm given here is advantageous if a few randomly selected combinations are needed and each selection is not based on the previous selection history.A one-to-one correspondence between the universe of n items and the first n natural numbers is established.The combinations produced by the algorithm are selections of p of the first n integers in lexicographical order [2, 5].Thus the combinations produced by the algorithm may be regarded equivalently as pointers to combinations of any objects.If the combinations are generated sequentially (according to the index), exactly the same order is achieved as that of Mifsud's algorithm [4].Vector generation is performed by probing the combination space and sequentially reducing the interval of uncertainty within which the indexed vector resides.Beginning with the leftmost vector position, trial values in ascending order for the digits are tested by computing the index of the (partial) trial vector, using essentially the method of Walter [7].The lexicographical index, i, for a combination