This paper modified the method of three-dimensional (3-dimensional) face reconstruction from a single two-dimensional (2-dimensional) image based on the Lambertian model consisting of height estimation, normal surface estimation, albedo calculation and image normalization, normal surface calculation, actual height calculation, and error correction. In height estimation, the facial height of each input image is estimated from the average facial height of face images in the training data set. The estimated height is used for estimating the normal surface by applying our proposed pattern morphing method (PMM). To calculate the actual normal surface, the albedo of the input face image is calculated to normalize the image first. Then, the actual normal surface is derived by using our proposed Y-ratio calculation with improved computational time. Finally, the actual height of input face image is computed afterwards to construct the 3-dimensional face. Two face databases of 110 human face images containing 10 real and 100 synthetic images were tested by our proposed method with uniform reflectance and high level of heterogeneous reflectance abilities. The experimental results were compared with the results obtained from other existing methods, such as the traditional minimization, shape propagation, local, and linear approaches. Our method can accurately reconstruct 3-dimensional face from a single 2-dimensional face image with the error less than 6%.

BACKGROUND: Human Papillomavirus (HPV) genotyping is an important approach to fight cervical cancer due to the relevant information regarding risk stratification for diagnosis and the better understanding of the relationship of HPV with carcinogenesis. This paper proposed two new feature extraction techniques, i.e. ChaosCentroid and ChaosFrequency, for predicting HPV genotypes associated with the cancer. The additional diversified 12 HPV genotypes, i.e. types 6, 11, 16, 18, 31, 33, 35, 45, 52, 53, 58, and 66, were studied in this paper. In our proposed techniques, a partitioned Chaos Game Representation (CGR) is deployed to represent HPV genomes. ChaosCentroid captures the structure of sequences in terms of centroid of each sub-region with Euclidean distances among the centroids and the center of CGR as the relations of all sub-regions. ChaosFrequency extracts the statistical distribution of mono-, di-, or higher order nucleotides along HPV genomes and forms a matrix of frequency of dots in each sub-region. For performance evaluation, four different types of classifiers, i.e. Multi-layer Perceptron, Radial Basis Function, K-Nearest Neighbor, and Fuzzy K-Nearest Neighbor Techniques were deployed, and our best results from each classifier were compared with the NCBI genotyping tool. RESULTS: The experimental results obtained by four different classifiers are in the same trend. ChaosCentroid gave considerably higher performance than ChaosFrequency when the input length is one but it was moderately lower than ChaosFrequency when the input length is two. Both proposed techniques yielded almost or exactly the best performance when the input length is more than three. But there is no significance between our proposed techniques and the comparative alignment method. CONCLUSIONS: Our proposed alignment-free and scale-independent method can successfully transform HPV genomes with 7,000 - 10,000 base pairs into features of 1 - 11 dimensions. This signifies that our ChaosCentroid and ChaosFrequency can be served as the effective feature extraction techniques for predicting the HPV genotypes.

Sport type classification and posture identification based on visual meaning of posture semantic in still images are challenging tasks. The difficulty of these tasks comes from the complex image content consisting of a player's posture, the color and texture of a player's clothes as well as complexity of the background. Player detection is one of the most important tasks in posture identification. For sport type classification without object segmentation, the new set of features, based on 64-bins color histogram, DCT coefficients, and Cb and Cr components, is introduced. To achieve high accuracy, an appropriate feature extraction technique should be also realized. For posture identification, three algorithms, concerning player region detection and suitable features for posture identification, are proposed namely blurred background elimination, irrelevant region elimination, and trimming players region. The DFT coefficients, based on image resizing and slicing techniques, are used as significant features in posture identification. Our proposed features were compared with Edge Histogram and Region-based Shape (EH and RS), two of MPEG-7 descriptors. The experimental results showed that our proposed features yielded better performance with 85.76% of accuracy in sport classification and 86.66% of accuracy in posture identification.

This paper proposes an unsupervised discrimination analysis for feature selection based on a property of the Fourier transform of the probability density distribution. Each feature is evaluated on the basis of a simple observation motivated by the concept of optical diffraction, which is invariant under feature scaling. The time complexity is O(mn), where m is number of features and n is number of instances when being applied directly to the given data. This approach is also extended to deal with data orientation, which is the direction of data alignment. Therefore, the discrimination score of any transformed space can be used for evaluating the original features. The experimental results on several real-world datasets demonstrate the effectiveness of the proposed method.

The challenge of clustering on data stream is the ability to deal with the continuous incoming data which are unlimited and unable to store all of them. To manage the storage crisis, the data must be processed in a single pass or only once after the arrival and are thrown away outer. All previously clustered data must be mathematically captured in terms of group features since those data are already non-existent. The proposed data stream clustering algorithm is divided into two main phases, namely on-line and off-line. In the on-line phase, new micro-cluster features are proposed. Our micro-cluster features better represent the arriving data than the traditional micro-cluster features. In the off-line phase, the prepared micro-clusters are categorized by their densities. The proposed method can generate the final clusters with different shapes and densities. Based on entropy, purity, Jaccard coefficient, and Rand statistic measures, our algorithm being applied on synthetic and real data outperforms the other previous data stream clustering algorithms.

In this paper, we propose a method to detect and identify commercial films from broadcast videos by using Temporal Occurrence Pattern (TOP). Our method uses the characteristic of broadcast videos in Japan that each individual commercial film appears multiple times in broadcast stream and typically has the same duration (e.g., 15 seconds). Using this characteristic, the method can detect as well as identify individual commercial films within given video archive. Based on simple signature (global feature) for each frame image, the method first puts all frames into numbers of buckets where each bucket contains frames having the same signature, and thus they appear the same. For each bucket, TOP as a binary sequence representing the occurrence time within video archive is then generated. All buckets are then clustered using simple hierarchical clustering with similarity between TOPs allowing possible temporal offset. This clustering stage can stitch up all frames for each commercial film and identify multiple occurrence of the same commercial film at the same time. We tested our method using actual broadcast video archive and confirmed good performance in detecting and identifying commercial films.

This paper proposes a very fast 1-pass-throw-away learning algorithm based on a hyperellipsoidal function that can be translated and rotated to cover the data set during learning process. The translation and rotation of hyperellipsoidal function depends upon the distribution of the data set. In addition, we present versatile elliptic basis function (VEBF) neural network with one hidden layer. The hidden layer is adaptively divided into subhidden layers according to the number of classes of the training data set. Each subhidden layer can be scaled by incrementing a new node to learn new samples during training process. The learning time is O(n), where n is the number of data. The network can independently learn any new incoming datum without involving the previously learned data. There is no need to store all the data in order to mix with the new incoming data during the learning process.

Singing voice recognition is a difficult topic in music information retrieval research area. The first approaches borrowed successful techniques widely used in automatic speech recognition (ASR) as speech and singing share similar acoustical feature since they are produced by the same apparatus. Moving from monophonic to polyphonic audio signal the problem become more complex as the background instrumental accompaniment is regarded as a noise source that has to be attenuated. This paper proposes a singing voice recognition algorithm that is able to automatically recognize the word in a singing signal with background music by using the concept of spectrogram pattern matching. The main idea is to apply both the spectrogram and the image processing methods to solve the problem of singing voice recognition. Each signal that accompanies music is analyzed and generated to its spectrogram that is used to train data for the classifier. Several classification functions are compared, such as Fisher classifier, feed-forward can effectively recognize the word in music with the accuracy rate more than 84%.

Existing clustering algorithms, such as single-link clustering, k-means, CURE, and CSM are designed to find clusters based on predefined parameters specified by users. These algorithms may be unsuccessful if the choice of parameters is inappropriate with respect to the data set being clustered. Most of these algorithms work very well for compact and hyper-spherical clusters. In this paper, a new hybrid clustering algorithm called Self-Partition and Self-Merging (SPSM) is proposed. The SPSM algorithm partitions the input data set into several subclusters in the first phase and, then, removes the noisy data in the second phase. In the third phase, the normal subclusters are continuously merged to form the larger clusters based on the inter-cluster distance and intra-cluster distance criteria. From the experimental results, the SPSM algorithm is very efficient to handle the noisy data set, and to cluster the data sets of arbitrary shapes of different density. Several examples for color image show the versatility of the proposed method and compare with results described in the literature for the same images. The computational complexity of the SPSM algorithm is O(N2), where N is the number of data points.

Invariant image recognition is one of the hardest problems in computer vision. The aim is to identify an image independently of its rotational orientation and size, as well as changing its color intensity. The current techniques such as high-ordered neural network and Zernike moments are not practical to apply to color images of size at least 256 /spl times/ 256 pixels. This paper introduces a novel recognition system, which has less computational time and space than the current techniques. Most implementations for finding machine classifiers are the multi layer perceptron (MLP) of which drawbacks lead to its costly training time and its slow learning behavior. The proposed system builds the classification function based on the concept of support vector machine (SVM) without kernel function of k-NN, called critical support vector machine (CSVM). CSVM is better generalization ability in terms of accuracy rate and lower complexity in terms of computational time and space. In part of invariance, the paper develops the rotational and scaling invariant self-organizing mapping (RSISOM), which composes the Kohonen's learning concept and principle component analysis. RSISOM can capture the features of color images invariant to rotation, scaling and color intensity. In experiment, the proposed system is tested against different 34 colored images with 98.2 % of the correct classification.

The paper presents a new 3D edge detector based on a cellular automata model. A uniform cellular automata rule using a 3D von Neumann neighborhood is proposed for carrying out edge detection on binary and gray-scaled 3D images. The work shows that a cellular automata-based model provides better edge maps for binary and gray-scaled objects when compared to the well-known 3D edge operators.

The problem of unsupervised classifying a set data and identifying the natural principal direction of each class at the same time is studied. A new adaptive unsupervised learning model called adaptive second order self-organizing map (ASOSOM) is proposed for this problem. ASOSOM combines the advantages of the self-organizing mapping with Karhunen-Loeve (KL) transformation. Instead of having one neuron representing each class, an additional neuron is introduced to cooperate with the class neuron for identifying the principal direction. Furthermore, a new performance measurement based on the co-variance between the natural principal direction and its perpendicular direction is introduced. This new model is applied to several applications and the obtained results are better than KL and MKL transformations.

Current unsupervised classification using self organizing mapping (SOM) competitive learning is based on the minimum Euclidean distance between a prototype neuron and the selected data. This is not suitable for several classification problems where the geometrical structure and curvature of the data space are the main concern. The problem studied in This work concerns the algorithm for measuring the non-Euclidean distance in a data point space, i.e. the surface function is unknown, and moving the prototype neurons along the actual geometrical structure of the data points. Our algorithm successfully classifies the experimental data spaces with various aspects while the SOM classification gives incorrect results.

Although several highly accurate blind source separation algorithms have already been proposed in the literature, these algorithms must store and process the whole data set which may be tremendous in some situations. This makes the blind source separation infeasible and not realisable on VLSI level, due to a large memory requirement and costly computation. This paper concerns the algorithms for solving the problem of tremendous data sets and high computational complexity, so that the algorithms could be run on-line and implementable on VLSI level with acceptable accuracy. Our approach is to partition the observed signals into several parts and to extract the partitioned observations with a simple activation function performing only the "shift-and-add" micro-operation. No division, multiplication and exponential operations are needed. Moreover, obtaining an optimal initial de-mixing weight matrix for speeding up the separating time will be also presented. The proposed algorithm is tested on some benchmarks available online. The experimental results show that our solution provides comparable efficiency with other approaches, but lower space and time complexity.

This article proposes an animating system for plant growth in the Lindenmayer (L) system in which alphabets are parameterized to the length, size, and position of each component of the plant. The growth rate is approximated by sigmoidal curves. To achieve a realistic leaf shape, a set of shape functions is used to transform a skeleton of primary branches into the final shape. Genetic algorithm (GA) is used to fit the shape's parameters. Modeling of the soybean is shown. The developments of plant growth are smooth and natural as well as realistic. This prototype can be used to generate a realistic model of any plant based on the L-system. © 2004 Wiley Periodicals, Inc.

Identifying the characteristics, structure, or features of a given set of data is the most essential step prior to the classification, recognition, or even understanding process. The identified features must preserve the nature of the data and the irrelevant features should be excluded after the extraction process. This study concerns the extraction of the actual 2-dimensional geometrical frame of each element from a set of given 2-dimensional structural objects. Our approach is based on self-organizing mapping and Karhunen-Loeve transformation. The number of clusters needs not be satisfied in advance as in the other approaches.

Invariant scaling and rotation recognition of an image has been successfully realized by extracting the features of the image based on various techniques such as moment, e.g., the Zernike moment, pulsed coupled neural network, and high order neural network. These approaches are costly in terms of computational time and network complexity. They are not practical when applied with an image of size at least 256 /spl times/ 256 pixels. In this paper, we reduce these complexities by applying the capability of a self-organizing mapping network such as Kohonen's competitive learning to extract the features. However, the competitive learning cannot be directly applied to this invariant scaling and rotation recognition problem. Some learning modifications are proposed so that no matter how an image is scaled or rotated the location of each neuron is always at the same coordinates with respect to its neighboring neurons. The new competitive learning was successfully tested with gray-scaled images.

Self-recovery micro-rollback synthesis (SMS) has currently become an important issue in high level synthesis. The problem of SMS combines the problem of functional unit scheduling and assignment with the problem of checkpoint insertion and microprogram optimization. It has been shown that these problems are NP-complete. The most studied problem is functional unit scheduling and assignment. Several heuristic techniques, including as soon as possible (ASAP), as last as possible (ALAP), integer programming, spring elasticity model, graph based mobility model, and genetic algorithm, are proposed. However, there are few studies on self-recovery micro-rollback synthesis and the technique of solution space searching by genetic algorithm has not been attempted. We study the feasibility of the genetic algorithm for the problem of SMS constrained on: the number of functional units, control steps, number of checkpoints, and the functional unit areas.

In this face recognition research, the eyebrows, eyes, nostrils, lips, and face contour are extracted separately. The shape, size, object-to-object distances, centroid, and orientation are found for each object. The techniques to solve the object shifting and rotating problems are investigated. Image subtraction is used to examine the geometric differences of the two different faces. The obtained classifying data in this research can accurately classify different people's faces.

The problem and efficient solution of automated synthesis of a self-recovery chip using micro-rollback and checkpoint insertion techniques are proposed and discussed. An efficient design of micro-rollback and checkpoint insertion can be achieved by considering them during the scheduling and allocation steps. The rollback and recovery scheme is designed to satisfy the constraints on the number of registers available and the maximum allowable recovery time. The proposed checkpointing (rollback point) algorithm will allow the system to recover from most transient faults.>

In this paper, a self-recovery technique of feedforward neural networks called weight shifting and its analytical models are proposed. The technique is applied to recover a network when some faulty links and/or neurons occur during the operation. If some input links of a specific neuron are detected faulty, their weights will be shifted to healthy links of the same neuron. On the other hand, if a faulty neuron is encountered, then we can treat it as a special case of faulty links by considering all the output links of that neuron to be faulty. The aim of this technique is to recover the network in a short time without any retraining and hardware repair. We also propose the hardware architecture for implementing this technique.

The problem and efficient solution of automated synthesis of a self-recovery chip using micro-roll-back and checkpoint insertion techniques is discussed and proposed. An efficient design of micro-roll-back and checkpoint insertion can be achieved by considering them during the scheduling and allocation steps. The rollback and recovery scheme is designed to satisfy the constraints of the available number of registers and the maximum allowable recovery time. The proposed checkpointing( rollback point) algorithm will allow the system to recover from most transient faults.

A new channel routing algorithm called DTR (Defect-Tolerant Routing) is investigated. This algorithm minimizes the total area and simultaneously maximizes the performance by reducing the critical area which can potentially be the source of logical faults caused by the bridging effects of spot defects. Experimental results show DTR produces less critical area than Yoshimura&Kuh's algorithm [1].

A comparison is made between two tree data structures that support region queries for VLSI layout systems. In a previous study, J.B. Rosenberg (ibid., vol.CAD-4, no.1, p.53-67, 1985) demonstrated that four-dimensional trees have a faster speed but use more memory space than quad trees with bisector lists. In this study the authors show that quad trees without bisector lists, for some small threshold values, have a better speed and require less storage than four-dimensional trees.>

The authors consider the constraint limiting multiple folding in programmable logic array (PLA) layouts imposed by the layout architecture which positions pull-up transistors on the boundary of the cell and uses another metal layer to connect pull-ups to terms inside the PLA. The general PLA architecture supports only input- and output-port folding by sharing them in either the same column or the same row. Term folding is allowed only with drastic changes in the layout architecture. Folding optimization algorithms, generally, have not considered a pull-up transistor placement as a constraint. A layout architecture is introduced and a technique for pull-up transistor folding based on a weighted-graph model is presented. The architecture supports also both I/O and term foldings. In comparison with other architectures the described architecture allows significant area improvement.>

The constraints limiting multiple PLA folding are the positions of pull-up transistors (Tpu) and the layout architecture which supports only I/O folding. In this paper, we present a novel architecture supporting multiple I/O, term, and pull-up foldings with through-the-cell net routing. The pull-up folding allows implementation of multiple level Boolean functions. Our architecture model achieves a better densities in comparison with PLA model.