Abstract In recent years, the dimensionality reduction has become more important as the number of dimensions of data used in various tasks such as regression and classification has increased. As popular nonlinear dimensionality reduction methods, t-distributed stochastic neighbor embedding (t-SNE) and uniform manifold approximation and projection (UMAP) have been proposed. However, the former outputs only one low-dimensional space determined by the t-distribution and the latter is difficult to control the distribution of distance between each pair of samples in low-dimensional space. To tackle these issues, we propose novel t-SNE and UMAP extended by q-Gaussian distribution, called q-Gaussian-distributed stochastic neighbor embedding (q-SNE) and q-Gaussian-distributed uniform manifold approximation and projection (q-UMAP). The q-Gaussian distribution is a probability distribution derived by maximizing the tsallis entropy by escort distribution with mean and variance, and a generalized version of Gaussian distribution with a hyperparameter q. Since the shape of the q-Gaussian distribution can be tuned smoothly by the hyperparameter q, q-SNE and q-UMAP can in- tuitively derive different embedding spaces. To show the quality of the proposed method, we compared the visualization of the low-dimensional embedding space and the classification accuracy by k-NN in the low-dimensional space. Empirical results on MNIST, COIL-20, OliverttiFaces and FashionMNIST demonstrate that the q-SNE and q-UMAP can derive better embedding spaces than t-SNE and UMAP.

Neural Radiance Fields (NeRF) have significantly impacted the novel view synthesis task. While it can render realistic images from unseen viewpoints, the boundaries of the object or scene in the areas rapidly changing the color and density are blurred. We present a novel NeRF regularization based on neighboring pixels’ differences to resolve this issue. This proposed regularized NeRF is named Neighboring Pixels Differences Regularized Neural Radiance Fields. Our key insight is that the pixel neighborhood relation in the image has essential information about the structure of the object and scene. Therefore, we propose to incorporate the neighborhood information into the original NeRF. We can easily implement the proposed method by adding a regularization term based on pixel neighborhood relation into the loss function of the vanilla NeRF. In the experiments, to evaluate the effectiveness of our method, we compared the proposed method with the original NeRF using synthetic and real-world multiview data. The experimental results showed that our method allowed for rendering high-quality novel views for both settings than the original NeRF.

Generative Adversarial Networks (GANs) can learn a highly informative latent space. By manipulating the latent representation in the latent space, we can control attributes in the generated images without modifying the model of GAN itself. Most existing studies find the direction vector of each semantic in the latent space and steer the latent vectors using linear computation to control semantic changes in the images. However, moving against the semantic direction still makes the latent representation subject to the entanglement of attributes in the latent space. Therefore, to solve the problem of attribute entanglement in latent space, we propose a decomposition method that decomposes the identity information subspace and the attribute information subspace from the pre-trained latent space and realize the semantic change of the latent representation by moving in the attribute information subspace. In our experiments, we evaluated its effectiveness and compared the images after the attribute change. The experiments demonstrated that our method can further solve the attribute entanglement problem in the latent space and effectively reduce the impact on other contents in the latent representation after changing the attributes.

We propose a simple and effective regularization method for instance segmentation, GLRDN-L2 (Graph Laplacian Regularization based on Differences of Neighboring Pixels). Instance segmentation is a challenging task in computer vision. For many years, ROI-based methods such as Mask R-CNN have dominantly presented the top performances; however, the recently proposed CondInst, which employs dynamic FCNs as a mask head and performs instance-aware mask prediction, outperforms Mask R-CNN. To our best knowledge, all methods optimize a model based on pixel-wise losses such as Dice Loss. Even with the results of high-resolution masks, there are problems such as blurred boundaries and hollows in the instances. We assume that these problems are due to the spatial structure and contextual information contained in the relationships between neighboring pixels not being incorporated well in the model. To address these problems, we propose a regularization that penalizes the errors in the spatial structure with a graph composed of the differences between neighboring pixels. We compare models trained with and without our regularization in CondInst to validate the effect of a natural extension that adds differentiation to the loss function and demonstrate performance improvement on both COCO and Cityscapes datasets.

For dimensionality reduction, t-distributed stochastic neighbor embedding (t-SNE) is famous. This technique represents the similarity between the pair of the samples in the high-dimensional space as Gaussian distribution. Then the similarity between the pair of the samples in the low-dimensional embedding space is also represented by using t-distribution to obtain the feature vectors in the embedding space from the high-dimensional data. The authors proposed q-Gaussian distributed stochastic neighbor embedding (q-SNE) as an extension of the t-SNE. The q-Gaussian distribution can express many distributions by setting hyperparameter q$= and includes the Gaussian distribution and the t-distribution as the special cases with hyperparameter q close to 1.0 and q = 2.0. However, these methods are applicable for a given data set and it is not possible to map new samples into the embedded space. To address this problem, the parametric t-SNE is proposed to construct the non-linear mapping by using a feed-forward neural network. In this paper, we propose a novel technique called parametric q-SNE with Convolutional Neural Network without pre-training. On MNIST, FashionMNIST, and COIL-20, the effectiveness of the parametric q -SNE is shown by using the visualization on 2-dimensional mapping, and the classification by using k nearest neighbors (k-NN) on the embedded space.

Deep learning has been successfully applied to a variety of tasks. However, a lot of unnecessary information for the task is included in the training data and it is difficult to automatically remove such unnecessary information in the trained model. For example, it is necessary to ignore the variations of the patients in the measured data for medical diagnosis. To address this problem, we propose a method that applies a module called Gradient Reversal Layer to train the model by removing unnecessary information and extracting only the relevant information for the target task. In this study, we conducted three experiments to demonstrate the effectiveness of the proposed method. The first one is to classify clothing images with unnecessary shift information and the second one is to identify the person from the face images with unnecessary facial expressions. Finally, the third one is to estimate the presence of diseases in medical data with unnecessary variations due to the differences of the patients. In all experiments, we obtained results where the unwanted information in the features was reduced, thereby improving the desired classification performance.

The dimensionality reduction has been widely introduced to use the high-dimensional data for regression, classification, feature analysis, and visualization. As the one technique of dimensionality reduction, a stochastic neighbor embedding (SNE) was introduced. The SNE leads powerful results to visualize high-dimensional data by considering the similarity between the local Gaussian distributions of high and low-dimensional space. To improve the SNE, a t-distributed stochastic neighbor embedding (t-SNE) was also introduced. To visualize high-dimensional data, the t-SNE leads to more powerful and flexible visualization on 2 or 3-dimensional mapping than the SNE by using a t-distribution as the distribution of low-dimensional data. Recently, Uniform manifold approximation and projection (UMAP) is proposed as a dimensionality reduction technique. We present a novel technique called a q-Gaussian distributed stochastic neighbor embedding (q-SNE). The q-SNE leads to more powerful and flexible visualization on 2 or 3-dimensional mapping than the t-SNE and the SNE by using a q-Gaussian distribution as the distribution of low-dimensional data. The q-Gaussian distribution includes the Gaussian distribution and the t-distribution as the special cases with q=1.0 and q=2.0. Therefore, the q-SNE can also express the t-SNE and the SNE by changing the parameter q, and this makes it possible to find the best visualization by choosing the parameter q. We show the performance of q-SNE as visualization on 2-dimensional mapping and classification by k-Nearest Neighbors (k-NN) classifier in embedded space compared with SNE, t-SNE, and UMAP by using the datasets MNIST, COIL-20, OlivettiFaces, FashionMNIST, and Glove.

Since the convolutional neural networks are often trained with redundant parameters, it is possible to reduce redundant kernels or filters to obtain a compact network without dropping the classification accuracy. In this paper, we propose a filter pruning method using the hierarchical group sparse regularization. It is shown in our previous work that the hierarchical group sparse regularization is effective in obtaining sparse networks in which filters connected to unnecessary channels are automatically close to zero. After training the convolutional neural network with the hierarchical group sparse regularization, the unnecessary filters are selected based on the increase of the classification loss of the randomly selected training samples to obtain a compact network. It is shown that the proposed method can reduce more than 50% parameters of ResNet for CIFAR-10 with only 0.3 % decrease in the accuracy of test samples. Also, 34% parameters of ResNet are reduced for TinyImageNet-200 with higher accuracy than the baseline network.

A deep convolutional neural network (CNN) has been widely used in image classification and gives better classification accuracy than the other techniques. The softmax cross-entropy loss function is often used for classification tasks. There are some works to introduce the additional terms in the objective function for training to make the features of the output layer more discriminative. The neuron-wise discriminant criterion makes the input feature of each neuron in the output layer discriminative by introducing the discriminant criterion to each of the features. Similarly, the center loss was introduced to the features before the softmax activation function for face recognition to make the deep features discriminative. The ReLU function is often used for the network as an active function in the hidden layers of the CNN. However, it is observed that the deep features trained by using the ReLU function are not discriminative enough and show elongated shapes. In this paper, we propose to use the neuron-wise discriminant criterion at the output layer and the center-loss at the hidden layer. Also, we introduce the online computation of the means of each class with the exponential forgetting. We named them adaptive neuron-wise discriminant criterion and adaptive center loss, respectively. The effectiveness of the integration of the adaptive neuron-wise discriminant criterion and the adaptive center loss is shown by the experiments with MNSIT, FashionMNIST, CIFAR10, CIFAR100, and STL10.

In a deep neural network (DNN), the number of the parameters is usually huge to get high learning performances. For that reason, it costs a lot of memory and substantial computational resources, and also causes overfitting. It is known that some parameters are redundant and can be removed from the network without decreasing performance. Many sparse regularization criteria have been proposed to solve this problem. In a convolutional neural network (CNN), group sparse regularizations are often used to remove unnecessary subsets of the weights, such as filters or channels. When we apply a group sparse regularization for the weights connected to a neuron as a group, each convolution filter is not treated as a target group in the regularization. In this paper, we introduce the concept of hierarchical grouping to solve this problem, and we propose several hierarchical group sparse regularization criteria for CNNs. Our proposed the hierarchical group sparse regularization can treat the weight for the input-neuron or the output-neuron as a group and convolutional filter as a group in the same group to prune the unnecessary subsets of weights. As a result, we can prune the weights more adequately depending on the structure of the network and the number of channels keeping high performance. In the experiment, we investigate the effectiveness of the proposed sparse regularizations through intensive comparison experiments on public datasets with several network architectures.

In recent years, deep learning has spread rapidly, and deeper, larger models have been proposed. However, the calculation cost becomes enormous as the size of the models becomes larger. Various techniques for compressing the size of the models have been proposed to improve performance while reducing computational costs. One of the methods to compress the size of the models is knowledge distillation (KD). Knowledge distillation is a technique for transferring knowledge of deep or ensemble models with many parameters (teacher model) to smaller shallow models (student model). Since the purpose of knowledge distillation is to increase the similarity between the teacher model and the student model, we propose to introduce the concept of metric learning into knowledge distillation to make the student model closer to the teacher model using pairs or triplets of the training samples. In metric learning, the researchers are developing the methods to build a model that can increase the similarity of outputs for similar samples. Metric learning aims at reducing the distance between similar and increasing the distance between dissimilar. The functionality of the metric learning to reduce the differences between similar outputs can be used for the knowledge distillation to reduce the differences between the outputs of the teacher model and the student model. Since the outputs of the teacher model for different objects are usually different, the student model needs to distinguish them. We think that metric learning can clarify the difference between the different outputs, and the performance of the student model could be improved. We have performed experiments to compare the proposed method with state-of-the-art knowledge distillation methods. The results show that the student model obtained by the proposed method gives higher performance than the conventional knowledge distillation methods.

In this work, we will consider the dimension reduction of the set of time series, such as economic data, to find the meaningful basis vector for the set of data, and indicate which data use which basis vector. Usually each of the time series is analyzed independently in economics but here we will analyze the set of time series simultaneously. Since some of the economic data are measured as positive values and we want to decompose them as a mixture of the parts, we will apply non-negative matrix factorization to the economic data. Non-negative matrix factorization can compress dimensions by approximating a non-negative matrix with the product of two non-negative matrices. The two non-negative matrices are called the coefficient matrix and the basis matrix, and the basis matrix can be considered as a dimensionally compressed matrix. If the standard non-negative matrix factorization is used for economic data, the basis matrix may not be smooth. We think that the basis vectors should be smooth except a few special economical incidents. In the proposed method, a Graph-based non-negative matrix factorization is introduced to regularize the basis matrix of the time series. A path graph for representing the time series of economic data is incorporated into the non-negative matrix factorization as regularization. As a result, basis vectors that maintains the time series of economic data are decomposed. Furthermore, we propose to introduce a sparsity in the non-negative matrix factorization. Traditionally, the sparsity incorporated into non-negative matrix factorization has been used for basis vectors. However, the proposed method introduces the sparsity for coefficient vectors. Thus the proposed method, which simultaneously incorporates the sparsity for the coefficient vectors and the smoothness for the basis vectors, can extract the smooth basis vectors and the original economical data are approximated as the weighted sum of the few bases vectors. This allows us to discover economic trends and the best-fit trends for each data at the same time.

Video Super resolution algorithms usually utilize the motion information of each pixel in consecutive frames to interpolate pixel values in a higher resolution and reconstruct frames of a higher resolution video from a lower resolution input video. Recently, architectures based on deep neural networks have gained popularity and can generate higher resolution videos with better visual quality. Also, deep neural networks make single image super resolution possible. In single image super resolution, a higher resolution image is constructed from a given input image by learning the transformation from lower resolution images to the higher resolution images. In this paper we propose to apply the single image super resolution algorithm to each frame in the original video and then use the resulting frames to estimate the movements of each pixel. Since the spatial resolution of the estimated motion of each pixel affects the visual quality of the higher resolution video, it is expected that the proposed approach can improve the visual quality of video super resolution. The proposed approach consistently results in good quality video reconstruction when tested on videos with diverse contents and different motion levels, which outperforms state of the art algorithms and offers competing visual performance on benchmark datasets.

This paper proposes an texture segmentation algorithm. In the proposed texture segmentation algorithm, the feature vectors at each pixel of an input image are extracted by using the deep neural networks such as the deep convolutional network (CNN) or the Siamese Network. Then they are used as input of the hierarchical region merging. Unlike the semantic segmentation such as fully connected network (FCN) or U-Net which are based on the supervised learning, the proposed algorithm can correctly segment the texture regions whose texture is taken from the other types of the texture. The effectiveness of the proposed texture segmentation algorithm is experimentally confirmed by using the famous texture images taken from book by P. Brodatz.

This paper introduces the sparse regularization for the convolutional neural network (CNN) with the rectified linear units (ReLU) in the hidden layers. By introducing the sparseness for the inputs of the ReLU, there is effect to push the inputs of the ReLU to zero in the learning process. Thus it is expected that the unnecessary increase of the outputs of the ReLU can be prevented. This is the similar effect with the Batch Normalization. Also the unnecessary negative values of the inputs of the ReLU can be reduced by introducing the sparseness. This can improve the generalization of the trained network. The relations between the proposed approach and the Batch Normalization or the modifications of the activation function such as Exponential Linear Unit (ELU) are also discussed. The effectiveness of the proposed method was confirmed through the detail experiments.

This paper describes a method for calibrating non-overlapping cameras in a simple way: using markers on the cameras. By adding an AR (Augmented Reality) marker to a camera, we can find the transformation between the fixed AR marker and the camera's center. With such information, relative pose of cameras can be easily found as long as the marker located on them is visible. Our method consists of the following two steps: (1) use an extra camera and a chessboard to find the transformation between the AR marker and the camera center. (2) Use the information from (1) and transformation between different markers to calibrate non-overlapping cameras. Compare to other non-overlapping calibration methods, our method can work with as less as one image and does not suffer from the degenerate cases.

Visual features trained from large scale image data by the deep convolutional neural network can be used for the other visual tasks. This paper investigates the effects of the learning of multiple tasks for such transfer learning from the source domains to the target domain. Two methods of the learning of multiple tasks are considered. Also we investigate which hidden layers should be re-trained for the target task in the fine-tuning process by selecting a subset of the hidden layers and updating only the parameters of the selected subset. Through a detail experiments, we confirmed the effectiveness of the learning of multiple tasks for pre-training. Also we showed that the first layer is not always required to be trained for the target task and the fully-connected layer, the classifier layer, and the last hidden layer should be retrained for the target task in the fine-tuning. These results suggests that the first and the second layers in the deep convolutional neural network trained by the learning of multiple tasks can extract general low level visual features.

We propose method for fast action recognition and comparable performance using local autocorrelation of optical flows over time. To capture action movement, dense optical flows is generated along sequence of video. Optical flows sometimes yield noise of motions that distract object of interest from another object motions and background. We suppress this by using edge based optical flow. The HOF vector is extracted from each window resolution and correlate its consecutive flow fields within cycle using local autocorrelation over time. It will gather richer information from movement while also gaining discriminative features than standard histogram methods. Comparison shows that the comparable performance is achieved over state of the arts.

There are two major approaches to content-based image retrieval using local image descriptors. One is descriptor-by-descriptor matching and the other is based on comparison of global image representation that describes the set of local descriptors of each image. In large-scale problems, the latter is preferred due to its smaller memory requirements; however, it tends to be inferior to the former in terms of retrieval accuracy. To achieve both low memory cost and high accuracy, we investigate an asymmetric approach in which the probability distribution of local descriptors is modeled for each individual database image while the local descriptors of a query are used as is. We adopt a mixture model of probabilistic principal component analysis. The model parameters constitute a global image representation to be stored in database. Then the likelihood function is employed to compute a matching score between each database image and a query. We also propose an algorithm to encode our image representation into more compact codes. Experimental results demonstrate that our method can represent each database image in less than several hundred bytes achieving higher retrieval accuracy than the state-of-the-art method using Fisher vectors.

In this paper, we propose a framework to assess visual complexity of paintings. This framework provides a machine learning scheme for investigating the relationship between human visual complexity perception and low-level image features. Since the global and local characteristics of paintings affect human's holistic impression and detail perception, we design a set of methods to extract the features that represent the global and local characteristics of paintings. By feature selection, we look into the role that each image feature plays in assessing visual complexity. Then the selected features are combined by a Support Vector Machine for classification. Experimental results indicate that the proposed work can predict the visual complexity perception of paintings with the accuracy of 88.13%, which is highly close to the assessments given by humans. Compared with the conventional measure of complexity, our approach considers human visual perception and performs more efficiently in assessing visual complexity of painting images.

Linear discriminant analysis (LDA) is a well-known method to extract efficient features for multi-class classification. Otsu derived the optimal (ultimate) non-linear discriminant analysis (ONDA) by supposing underlying probabilities and showed that ONDA was closely related to Bayesian decision theory (posterior probabilities). Also Otsu pointed out that the usual LDA could be regarded as the linear approximation of this ultimate ONDA through the linear approximations of the Bayesian posterior probabilities. This theory of ONDA suggests that we can construct a novel nonlinear discriminant mapping by utilizing the estimates of the posterior probabilities. Based on this theory, logistic discriminant analysis (LgDA) was proposed by one of the authors as the approximation of ONDA. In LgDA, the posterior probabilities are estimated by logistic regression. In this paper, we propose the sparse logistic discriminant analysis in which the posterior probabilities are estimated by the sparse logistic regression with L2-or L1-regularizer to improve the generalization performance of LgDA further. Experiments using the standard datasets for classification reveal that the discriminant spaces by our proposed method (LgDA-L2 and LgDA-L1) are better than those by LDA and LgDA in terms of the recognition rates for test samples.

This study aimed to realize and compare the effectiveness of cortical and trabecular bone measures in a computer-aided system based on support vector machine (SVM) for identifying postmenopausal women with low BMD or osteoporosis. Dental panoramic radiographs of 69 postmenopausal women, as well as bone mineral density (BMD) assessments at the lumbar spine and the femoral neck were used. Average width and variance of continuous measurements of mandibular cortical bone and distribution of average length and angle (direction) of trabecular bone segments on panoramic radiographs were measured by computer-aided systems and used as inputs. The accuracy of the cortical measures using average width and variance was 87% and of the trabecular measures using average length and angle was 65% with RBF kernel-SVM method for diagnosing women with low BMD at the lumbar spine. Likewise, the accuracy of combined cortical and trabecular bone measures was 88%. Our results suggest that the cortical bone measurement is more appropriate than the trabecular bone measurement for triage screening of osteoporosis.

Linear discriminant analysis (LDA) is one of the well known methods to extract the best features for multi-class discrimination. Recently Kernel discriminant analysis (KDA) has been successfully applied in many applications. KDA is one of the nonlinear extensions of LDA and construct nonlinear discriminant mapping by using kernel functions. But the kernel function is usually defined a priori and it is not known what the optimum kernel function for nonlinear discriminant analysis is. Also the class information is not usually introduced to define the kernel functions. In this paper the optimum kernel function in terms of the discriminant criterion is derived by investigating the optimum discriminant mapping constructed by the optimum nonlinear discriminant analysis (ONDA). Otsu derived the optimum nonlinear discriminant analysis (ONDA) by assuming the underlying probabilities similar with the Bayesian decision theory. He showed that the optimum non linear discriminant mapping was obtained by using Variational Calculus. The optimum nonlinear discriminant mapping can be defined as a linear combination of the Bayesian a posterior probabilities and the coefficients of the linear combination are obtained by solving the eigenvalue problem of the matrices defined by using the Bayesian a posterior probabilities. This means that the ONDA is closely related to Bayesian decision theory. Also Otsu showed that LDA could be interpreted as a linear approximation of the ONDA through the linear approximation of the Bayesian a posterior probabilities. In this paper, the optimum kernel function is derived by investigating the optimum discriminant mapping constructed by ONDA. The derived kernel function is also given by using the Bayesian a posterior probabilities. This means that the class information is naturally introduced in the kernel function. For real application, we can define a family of discriminate kernel functions can be defined by changing the estimation method of the Bayesian a posterior probabilities.

This paper proposes a spatio-temporal feature using three-way cross-correlations of local motion attributes for action recognition. Recently, the cubic higher-order local auto-correlations (CHLAC) feature has been shown high classification performances for action recognition. In previous researches, CHLAC feature was applied to binary motion image sequences that indicates moving or static points. However, each binary motion image lost informations about the type of motion such as timing of change or motion direction. Therefore, we can improve the classification accuracy further by extending CHLAC to multivalued motion image sequences that considered several types of local motion attributes. The proposed method is also viewed as an extension of popular bag-of-features approach. Experimental results using two datasets shows proposed method outperformed CHLAC features and bag-of-features approach.

A novel visual tracking algorithm is proposed in this paper. The algorithm uses pixel-pair features to discriminate between an image patch with an object in the correct position and image patches with an object in an incorrect position. The pixel-pair feature is considered to be robust for the illumination change, and also is robust for partial occlusion when appropriate features are selected in every video frame. The tracking precision for a deforming object (skier) is examined and also the occlusion detection method is described.

Linear discriminant analysis (LDA) is one of the well known methods to extract the best features for the multi-class discrimination. Otsu derived the optimal nonlinear discriminant analysis (ONDA) by assuming the underlying probabilities and showed that the ONDA was closely related to Bayesian decision theory (the posterior probabilities). Also Otsu pointed out that LDA could be regarded as a linear approximation of the ONDA through the linear approximations of the Bayesian posterior probabilities. Based on this theory, we propose a novel nonlinear discriminant analysis named logistic discriminant analysis (LgDA) in which the posterior probabilities are estimated by multi-nominal logistic regression (MLR). The experimental results are shown by comparing the discriminant spaces constructed by LgDA and LDA for the standard repository datasets.

Adaboost is an ensemble learning algorithm that combines many base-classifiers to improve their performance. Starting with Viola and Jones’ researches, Adaboost has often been used to local feature selection for object detection. Adaboost by Viola-Jones consists of following two optimization schemes: (1) training of the local features to make base-classifiers, and (2) selection of the best local feature. Because the number of local features becomes usually more than tens of thousands, the learning algorithm is time consuming if the two optimizations are completely performed. To omit the unnecessary redundancy of the learning, we propose fast boosting algorithms by using Particle Swarm Optimization (PSO) and random candidate selection (RCS). Proposed learning algorithm is 50 times faster than the usual Adaboost while keeping comparable classification accuracy.

An example-based classification algorithm to improve generalization performance for detecting objects in images is presented. The classifier integrates component-based classifiers according to the AdaBoost algorithm. A probability estimate by a kernel-SVM is used for the outputs of base learners, which are independently trained for local features. The base learners are determined by selecting the optimal local feature according to sample weights determined by the boosting algorithm with cross-validation. Our method was applied to the MIT CBCL pedestrian image database, and 54 sub-regions were extracted from each image as local features. The experimental results showed a good classification ratio for unlearned samples.

We propose an automatic factorization method for time series signals that follow Boltzmann distribution. Generally time series signals are fitted by using a model function for each sample. To analyze many samples automatically, we have to apply a factorization method. When the energy dynamics are measured in thermal equilibrium, the energy distribution can be modeled by Boltzmann distribution law. The measured signals are factorized as the non-negative sum of the probability density function of Boltzmann distribution. If these signals are composed from several components, then they can be decomposed by using the idea of non-negative matrix factorization (NMF). In this paper, we modify the original NMF to introduce the probability density function modeled by Boltzmann distribution. Also the number of components in samples is estimated by using model selection method. We applied our proposed method to actual data that was measured by fluorescence correlation spectroscopy (FCS). The experimental results show that our method can automatically factorize the signals into the correct components.

Kernel logistic regression (KLR) is a powerful discriminative algorithm. It has similar loss function and algorithmic structure to the kernel support vector machine (SVM). Recently, Zhu and Hastie proposed the import vector machine (IVM) in which a subset of the input vectors of KLR are selected by minimizing the regularized negative log-likelihood to improve the generalization performance and to reduce computation cost. In this paper, two modifications of the original IVM are proposed. The cross-validation based criterion is used to select import vectors instead of the likelihood based criterion. Also binary particle swarm optimization is used to select good subset instead of the greedy stepwise algorithm of the original IVM. Through the comparison experiment, the improvement of the generalization performance of the proposed algorithm was confirmed.

BACKGROUND: Recent analyses have suggested that many genes possess multiple transcription start sites (TSSs) that are differentially utilized in different tissues and cell lines. We have identified a huge number of TSSs mapped onto the mouse genome using the cap analysis of gene expression (CAGE) method. The standard hierarchical clustering algorithm, which gives us easily understandable graphical tree images, has difficulties in processing such huge amounts of TSS data and a better method to calculate and display the results is needed. RESULTS: We use a combination of hierarchical and non-hierarchical clustering to cluster expression profiles of TSSs based on a large amount of CAGE data to profit from the best of both methods. We processed the genome-wide expression data, including 159,075 TSSs derived from 127 RNA samples of various organs of mouse, and succeeded in categorizing them into 70-100 clusters. The clusters exhibited intriguing biological features: a cluster supergroup with a ubiquitous expression profile, tissue-specific patterns, a distinct distribution of non-coding RNA and functional TSS groups. CONCLUSION: Our approach succeeded in greatly reducing the calculation cost, and is an appropriate solution for analyzing large-scale TSS usage data.

We consider the problem of recognizing a specific human face in different poses (viewing direction) when only one frontal face image exists in the face database. To solve this problem, prior knowledge is learned by using principal component analysis on a set of multi-view images to obtain aligned principal components. They are used together with the idea of linear object classes to synthesize a virtual view of the frontal face from a given face image taken from a different viewing direction. The estimated virtual frontal view is then compared with the stored frontal face images in the face database to identify the person. Experimental results are shown using face images captured from different viewpoints.

Face tracking continues to be an important topic in computer vision. We describe a tracking algorithm based on a static face detector. Our face detector is a rectanglefeature- based boosted classifier, which outputs the confidence whether an input image is a face. The function that outputs this confidence, called a score function, contains important information about the location of a moving target. A target that has moved will be located in the gradient direction of a score function from the location before moving. Therefore, our tracker will go to the region where the score is maximum using gradient information of this function. We show that this algorithm works by the combination of jumping to the gradient direction and precise search at the local region.

The paper describes how to improve the robustness to occlusions in face recognition and detection. We propose a neural network architecture which integrates an auto-associative neural network into a simple classifier. The auto-associative network is employed to recall the original face from a partially occluded face image and to detect the occluded regions in the input image. The original face can be reconstructed by replacing those regions with the recalled pixels. By applying this reconstruction process recursively, the integrated network is able to classify occluded faces robustly. To confirm the effectiveness of this method, we performed experiments on face image classification and face detection. It is shown that the classification performance is not decreased even if 20-30% of the face image is occluded.

This paper introduces a kernel logit approach for face and non-face classification. The approach is based on the combined use of the multinomial logit model (MLM) and "kernel feature compound vectors." The MLM is one of the neural network models for multiclass pattern classification, and is supposed to be equal or better in classification performance than linear classification methods. The "kernel feature compound vectors" are compound feature vectors of geometric image features and Kernel features. Evaluation and comparison experiments were conducted by using face and non,face images (Face: training 100, cross-validation 300, test 325, Non-face : training 200, cross-validation 1000, test 1000) gathered from the available face databases and others. The experimental result obtained by the proposed method was better than the results obtained by the Support Vector Machines (SVM) and the Kernel Fisher Discriminant Analysis (KFDA).

This paper presents a face matching method through information theoretical attention points. The attention points are selected as the points where the outputs of Gabor filters applied to the contrast-filtered image (Gabor features) have rich information. The information value of Gabor features of the certain point is used as the weight and the weighed sum of the correlations is used as the similarity measure for the matching. To cope with the scale changes of a face, several images with different scales are generated by interpolation from the input image and the best match is searched. By using the attention points given from the information theoretical point of view, the matching becomes robust under various environments. This matching method is applied to face detection of a known person and face classification. The effectiveness of the proposed method is confirmed by experiments using the face images captured over years under the different environments.

This paper describes an image compression system based on evolvable hardware (EHW) for high precision printers (HPP). These printers are especially flexible for book publishing, but require large disk space for images, in particular those of higher resolution. To increase the printing speed and reduce the disk space, the images should be compressed. The system for this compression must be (1) adaptive, so that it changes depending on image characteristics and (2) on-line, which means implemented in hardware. The standard compression methods have a simple template change strategy which is not efficient for the images of HPP. We used an EHW system for compressing HPP images in real time. The EHW is a type of adaptive hardware which allows evolutionary algorithms to change the hardware configuration in real time. It works as fast as other compression systems (like the JBIG standard), but changes the image modeling to reflect the changes in the image characteristics. Simulation results show more than a 50% increase in compression ratio compared to JBIG for the printer system.

A basic color block truncation coding (CBTC) algorithm for color image compression is described. A modification of the algorithm that reduces truncation errors is also described. The block statistics related to CBTC methods are investigated. Some experimental results are given for a 256-*256-pixel color image with 24 b/pixel.>

Gives a basic idea and its fundamental algorithms of the visual interface for image database systems. The QVE (Query by Visual Example) accepts a sketch roughly drawn by a user to retrieve the original image and the similar images. The system evaluates the similarity between the rough sketch, i.e. a visual example, and each of the image data in the database automatically. The QVE interface is implemented and examined on an experimental electronic art gallery called ART MUSEUM. This paper also gives some experimental results and a current evaluation. The algorithms are quite effective for content based image retrieval.>

Proposes a face recognition method which is characterized by structural simplicity, trainability and high speed. The method consists of two stages of feature extractions: first, higher order local autocorrelation features which are shift-invariant and additive are extracted from an input image; then those features are linearly combined on the basis of multivariate analysis methods so as to provide new effective features for face recognition in learning from examples.>

A complex autoregressive model for invariant feature extraction to recognize arbitrary shapes on a plane is presented. A fast algorithm to calculate complex autoregressive coefficients and complex PARCOR coefficients of the model is also shown. The coefficients are invariant to rotation around the origin and to choice of the starting point in tracing a boundary. It is possible to make them invariant to scale and translation. Experimental results that the complicated shapes like nonconvex boundaries can be recognized in high accuracy, even in the low-order model. It is seen that the complex PARCOR coefficients tend to provide more accurate classification than the complex AR coefficients.>