Search

Recently, many deep learning methods are applied with the spatial information to learn features for polarimetric synthetic aperture radar (PolSAR) image classification. However, without considering the multi-scale information, the classification performance of these methods are limited. Hence, this paper proposes a multi-scale feature extraction method based on stacked sparse autoencoder (SSAE), named the multi-scale SSAE (MS-SSAE), to improve the classification performance. This method extracts the deep multi-scale features by a two-stage framework. In the first stage, the SSAE uses training data at different scales to extract the multi-scale features. Then, a 1-D average pooling strategy is proposed to reduce the feature dimensionality at the second stage. Therefore, the MS-SSAE can capture discriminative multi-scale features. The experimental results certify that the MS-SSAE can not only improve the classification accuracy, but also remain the details in the image. [ABSTRACT FROM AUTHOR]

Abstract In this paper, a fast unsupervised deep fusion framework for change detection of multitemporal synthetic aperture radar (SAR) images is presented. It mainly aim at generating a difference image (DI) in the feature learning procedure by stacked auto-encoders (SAEs). Stacked auto-encoders, as one kind of deep neural network, can learn feature maps that retain the structural information but suppress the noise in the SAR images, which will be beneficial for DI generation. Compared with shallow network, the proposed framework can extract more available features, and be favorable for getting better change results. Different with other common deep neural networks, our proposed method does not need labeled data to train the network. In addition, we find a subset of the entire samples that appropriately represent the whole dataset to speed up the training of the deep neural network without under-fitting. Moreover, we design a fusion network structure that can combine ratio operator based method to ensure that the representations of higher layers are better than that of the lower ones. To summarize, the main contribution of our work lies in using of deep fusion network for generation of DI in a fast and unsupervised way. Experiments on four real SAR images confirm that our network performs better than traditional ratio methods and convolutional neural network. [ABSTRACT FROM AUTHOR]

Inspired by the compressive sensing (CS) theory, a new classifier based on compressed dictionary and Least Squares Support Vector Machine (LS-SVM) is proposed to deal with large scale problems. The coefficients of support vectors can be recovered from a few measurements if LS-SVM is approximated to sparse structure. Using the known Cholesky decomposition, we approximate the given kernel matrix to represent the coefficients of support vectors sparsely by a low-rank matrix that we have used as a dictionary. The proposed measurement matrix being coupled with the dictionary forms a compressed dictionary that proves to satisfy the restricted isometry property (RIP). Our classifier has the quality of low storage and computational complexity, high degree of sparsity and information preservation. Experiments on benchmark data sets show that our classifier has positive performance. [ABSTRACT FROM AUTHOR]

It remains one of the most challenging tasks to distinguish different terrain materials from a single SAR image. With the increase of ground resolution, it allows us to model the SAR image directly by exploiting spatial structures and texture information that are extracted by several machine learning approaches. In this paper, a novel feature learning approach is proposed to capture discriminant features of high-resolution SAR images. In the first stage, a weighted discriminant filter bank is learned from some labeled SAR image patches to generate low-level features. Then, the locality constraint is introduced to produce the high-level features in both the encoding and the spatial pooling procedure. In this work, the superpixels are employed as the basic operational units instead of the pixels for terrain classification. With some learned domain patterns which are learned from all of the high-level features of each pixel, the superpixel is characterized by a hyper-feature. In the last stage, a linear-kernel support vector machine is utilized to classify all of these hyper-features which are generated for each superpixel. The experimental results show a better classification performance of the proposed approach than several available state-of-the-art approaches. [ABSTRACT FROM AUTHOR]

Many works have been recently presented to extract efficient features for automatic target recognition of synthetic aperture radar (SAR) images. However, they are limited in the discriminative ability of similar targets and robustness to the remarkable speckle noises and background clutters existed in images. In this paper, we propose a Complementary Spatial Pyramid Coding (CSPC) approach in the framework of Spatial Pyramid Matching (SPM). Both the coding coefficients and coding residuals are explored to develop more discriminative and robust features for representing SAR images. Multiple codebooks are first built from some training example images, where each codebook is formulated by local features of a certain class of samples. Then multiple sparse coding models are developed to derive features of a target under these codebooks. Additionally, these coding residuals are further sparsely encoded in the same way to that of local features. Finally, the encoded local features and the residual features are pooled according to spatial pyramid respectively, then concatenated to form the complementary features for the subsequent classification. The experiments on Moving and Stationary Target Acquisition and Recognition (MSTAR) public database verify the superior performance of the proposed method to some related approaches. [ABSTRACT FROM AUTHOR]

Spectral correlation between each band of multispectral (MS) image is an important characteristic. In this paper, a novel pan-sharpening method inspired by distributed compressed sensing (DCS) theory is presented, which considers the correlation characteristic among the MS bands as prior information in the restoration model. Two basic image formation models reflecting the relationships of the low-resolution multispectral (LRMS) image and the high-resolution panchromatic (HRP) image to the unknown high-resolution multispectral (HRMS) image are constructed. In order to exactly recover the HRMS image from the measurements, a joint sparsity model (JSM) is employed to solve the ill-posed inverse problem. The basic assumption of JSM is that each HRMS spectral band shares a sparse common component and a sparse innovation component. The choice of dictionaries that are used to sparsely represent the common and innovation components is very important. The common component can be sparsely coded by a dictionary, which is learned from the source HRP image patches. Each innovation component can be sparsely represented by a hybrid dictionary, which is composed of discrete cosine transform (DCT) bases, Gabor bases, wavelet ‘db1’ bases, ridgelet bases and the learned dictionary atoms. By solving the l 1 -norm optimization problem, the unknown HRMS image can be estimated. The proposed method and the state-of-the-art methods are performed on simulated and real remote sensing images. Experimental results demonstrate that the proposed pan-sharpening method shows better performance than other well-known methods in terms of quantitative assessment and visual analysis. [ABSTRACT FROM AUTHOR]

Abstract: Low rank representation is capable of capturing the global structure of mixed subspaces which are usually assumed to be independent. However, its computation is time-consuming. In practice, the data always distributes on subspaces that intersect or even overlap with each other. So the local structure of the data among the overlapping parts is important. Sparsity is a good property to accelerate the algorithm and capture the local linear structure. The drawback is that it breakups the low rank property of the reconstruction coefficient matrix when combining with low rank representation. In order to combine the two advantages properly, in this paper, we introduce a new constraint to the low rank representation matrix, which is called sparse congruency. Fortunately, we find that this new constraint is a simplification to the low rank and sparse constraints. Several experiments are implemented to demonstrate the efficiency of our method in semi-supervised classification. [Copyright &y& Elsevier]

Abstract: A hybrid memetic algorithm, called a memetic algorithm with double mutation operators (MADM), is proposed to deal with the problem of global optimization. In this paper, the algorithm combines two meta-learning systems to improve the ability of global and local exploration. The double mutation operators in our algorithms guide the local learning operator to search the global optimum; meanwhile the main aim is to use the favorable information of each individual to reinforce the exploitation with the help of two meta-learning systems. Crossover operator and elitism selection operator are incorporated into MADM to further enhance the ability of global exploration. In the first part of the experiments, six benchmark problems and six CEC2005׳s problems are used to test the performance of MADM. For the most problems, the experimental results demonstrate that MADM is more effective and efficient than other improved evolutionary algorithms for numerical optimization problems. In the second part of the experiments, MADM is applied to a practical problem, clustering complex and linearly non-separable datasets, with a satisfying result. [Copyright &y& Elsevier]

We present an Augmentative Contrastive Learning for One-Shot Object Detection method that is inspired by the co-attention and co-excitation (CoAE) method. In the One-shot Object Detection (OSOD) task, the target image contains rich background information that influences the outcomes of different categories of objects more significantly. As a result, the network frequently suggests objects that do not fall under this category, leading to false detections. To learn the similarity between the target object and the query patch, we built the data augmentation module of object switch by similarity, encourage and restrain module. It primarily consists of the following three components: 1) To improve the contrast between the foreground and background, we propose an object switch module based on data augmentation and feature similarity; 2) To decrease the likelihood of false detections, we propose an encourage and restrain module that allows the network to encourage the recommendation of similar objects while restricting the recommendation of different categories of objects; and 3) To increase classification results, we use the nonlinear projector. We perform tests on the PASCAL VOC and MS COCO datasets to confirm the efficacy of our technique. [ABSTRACT FROM AUTHOR]

Abstract: Ng–Jordan–Weiss (NJW) method is one of the most widely used spectral clustering algorithms. For a K clustering problem, this method partitions data using the largest K eigenvectors of the normalized affinity matrix derived from the dataset. It has been demonstrated that the spectral relaxation solution of K-way grouping is located on the subspace of the largest K eigenvectors. However, we find from a lot of experiments that the top K eigenvectors cannot always detect the structure of the data for real pattern recognition problems. So it is necessary to select eigenvectors for spectral clustering. We propose an eigenvector selection method based on entropy ranking for spectral clustering (ESBER). In this method, first all the eigenvectors are ranked according to their importance on clustering, and then a suitable eigenvector combination is obtained from the ranking list. In this paper, we propose two strategies to select eigenvectors in the ranking list of eigenvectors. One is directly adopting the first K eigenvectors in the ranking list. Different to the largest K eigenvectors of NJW method, these K eigenvectors are the most important eigenvectors among all the eigenvectors. The other eigenvector selection strategy is to search a suitable eigenvector combination among the first Km (Km>K) eigenvectors in the ranking list. The eigenvector combination obtained by this strategy can reflect the structure of the original data and lead to a satisfying spectral clustering result. Furthermore, we also present computational complexity reduction strategies for ESBER method to deal with large-scale datasets. We have performed experiments on UCI benchmark datasets, MNIST handwritten digits datasets, and Brodatz texture datasets, adopting NJW method for a baseline comparison. The experimental results show that ESBER method is more robust than NJW method. Especially, ESBER method with the latter eigenvector selection strategy can obtain satisfying clustering results in most cases. [Copyright &y& Elsevier]

Abstract: Synaptic transmission between neocortical neurons often shows activity-dependent synaptic depression. Moreover, it is revealed recently that the synaptic depression has a more active role in information processing. The population density approach is a viable method to describe the large populations of neurons and has generated considerable interest recently. In this paper, we use the population stochastic dynamics to analyze the population of neurons with synaptic depression in the stochastic environment. We have derived an evolution equation of the membrane potential density function with synaptic depression, and obtain several formulas for analytic computing the response of instantaneous fire rate. Through a technical analysis, we arrive at several significant conclusions, which show that the background inputs and the spatial distribution of synapses and the spatial–temporal relation of inputs play an important role in information processing. [Copyright &y& Elsevier]

Abstract: Many immue-inspired algorithms are based on the abstractions of one or several immunology theories, such as clonal selection, negative selection, positive selection, rather than the whole process of immune response to solve computational problems. In order to build a general computational framework by simulating immune response process, this paper introduces a model for population-based artificial immune systems, termed as PAIS, and applies it to numerical optimization problems. PAIS models the dynamic process of human immune response as a quaternion ( G , I , R , Al ), where G denotes exterior stimulus or antigen, I denotes the set of valid antibodies, R denotes the set of reaction rules describing the interactions between antibodies, and Al denotes the dynamic algorithm describing how the reaction rules are applied to antibody population. Some general descriptions of reaction rules including the set of clonal selection rules and the set of immune memory rules are introduced in PAIS. Based on these reaction rules, a dynamic algorithm, termed as PAISA, is designed for numerical optimization. In order to validate the performance of PAISA, nine benchmark functions with 20–10,000 dimensions and a practical optimization problem, optimal approximation of linear systems are solved by PAISA, successively. The experimental results indicate that PAISA has high performance in optimizing some benchmark functions and practical optimization problems. [Copyright &y& Elsevier]

Traditional unsupervised feature selection methods usually construct a fixed similarity matrix. This matrix is sensitive to noise and becomes unreliable, which affects the performance of feature selection. The researches have shown that both the global reconstruction information and local structure information are important for feature selection. To solve the above problem effectively and make use of the global and local information of data simultaneously, a novel algorithm is proposed in this paper, called subspace learning for unsupervised feature selection via adaptive structure learning and rank approximation (SLASR). Specifically, SLASR learns the manifold structure adaptively, thus the preserved local geometric structure can be more accurate and more robust to noise. As a result, the learning of the similarity matrix and the low-dimensional embedding is completed in one step, which improves the effect of feature selection. Meanwhile, SLASR adopts the matrix factorization subspace learning framework. By minimizing the reconstruction error of subspace learning residual matrix, the global reconstruction information of data is preserved. Then, to guarantee more accurate manifold structure of the similarity matrix, a rank constraint is used to constrain the Laplacian matrix. Additionally, the l 2,1/2 regularization term is used to constrain the projection matrix to select the most sparse and robust features. Experimental results on twelve benchmark datasets show that SLASR is superior to the six comparison algorithms from the literature. [ABSTRACT FROM AUTHOR]

Cross-Domain Few-Shot Learning (CD-FSL) to recognize new categories in a new domain with few samples has attracted significant attention. Recently, task-specific CD-FSL emerges as promising research for its great generalization. However, the existing task-specific CD-FSL methods are not robust enough in task context modeling and task-specific feature learning, especially under the cross-domain setting. To tackle this problem, a Task Context Transformer and Graph Convolutional Network (TCT-GCN) method for CD-FSL is proposed. The proposed TCT-GCN constructs three modules: 1) Multi-Level Feature Fusion, which fuses domain-shared low-level features into domain-unshared high-level features; 2) Transformer-based Task Context Encoder, which models sample order-independent task context; 3) Graph Convolutional Network-based Adaptive Feature Learner, which adaptively learns task-specific features. Experiments on eight CD-FSL datasets reveal the effectiveness of our TCT-GCN. [ABSTRACT FROM AUTHOR]

Recently, a massive number of deep learning-based approaches have been successfully applied to various remote sensing image (RSI) recognition tasks. However, most existing advances of deep learning methods in the RSI field heavily rely on the features extracted by the manually designed backbone network, which severely hinders the potential of deep learning models due to the complexity of RSI and the limitation of prior knowledge. In this paper, we research a new design paradigm for the backbone architecture in RSI recognition tasks, including scene classification, land-cover classification and object detection. A novel one-shot architecture search framework based on a weight-sharing strategy and an evolutionary algorithm is proposed, called RSBNet, which consists of three stages: Firstly, a supernet constructed in a layer-wise search space is pretrained on a self-assembled large-scale RSI dataset based on an ensemble single-path training strategy. Next, the pre-trained supernet is equipped with different recognition heads through the switchable recognition module and respectively fine-tuned on the target dataset to obtain task-specific supernet. Finally, we search for the optimal backbone architecture for different recognition tasks based on the evolutionary algorithm without any network training. Extensive experiments have been conducted on five benchmark datasets for different recognition tasks, the results show the effectiveness of the proposed search paradigm and demonstrate that the searched backbone is able to flexibly adapt different RSI recognition tasks and achieve impressive performance. [ABSTRACT FROM AUTHOR]

Many traditional unsupervised feature selection algorithms utilize manifold information to mine the local structure of the data. However, the noise existing in the raw data reduces the accuracy of the manifold information of data, which affects the learning effect of the entire algorithm. In order to solve the above problems and more fully find the internal structure inside the data, this paper proposes an adaptive graph regularization and self-expression for noise-aware feature selection (ASNFS). Firstly, the algorithm adopts non-negative matrix factorization to decompose the raw data matrix, and utilizes the low-dimensional matrix generated after decomposition to replace the raw high-dimensional data matrix. This allows the algorithm to reveal some internal structural information of the raw data while reducing the dimensionality of the data. ASNFS also introduces the orthogonal basis clustering with excellent clustering effect, and the interpretability of the algorithm is enhanced. Secondly, in addition to preserving the manifold information in the low-dimensional projection subspace, the algorithm also preserves the manifold information in the non-negative matrix factorization subspace. Meanwhile, the adaptive graph regularization term added to the objective function enables the algorithm to continuously update the similarity matrix. It can effectively remove the noise inside the raw data, and prevent the occurrence of over-fitting phenomenon of the experimental results caused by the fixed similarity matrix. Finally, the similarity matrix retains the local structure information of the data with each iteration, and the results of the feature selection are reused for the construction of the similarity matrix. ASNFS adopts an alternate iterative method to optimize the objective function, which is simple and effective. Then, the algorithm complexity and convergence are analyzed. ASNFS is compared with seven feature selection algorithms on nine datasets, and the experimental results reflect the effectiveness of ASNFS in feature selection. [ABSTRACT FROM AUTHOR]

Non-negative matrix factorization (NMF) can map high-dimensional data into a low-dimensional data space. Feature selection can eliminate the redundant and irrelevant features from the alternative features. In this paper, we propose a feature selection based dual-graph sparse non-negative matrix factorization (DSNMF) which can find an appropriate low dimensional representation of data by NMF and then select more discriminative features to further reduce the dimension of the low dimensional space by feature selection rather than reduce the dimension by only NMF or feature selection in many previous methods. DSNMF combines dual-graph model with non-negative matrix factorization, which can not only simultaneously preserve the geometric structures in both the data space and the feature space, but also make the two non-negative matrix factors update iteratively and interactively. In addition, DSNMF exerts L 2,1 - norm constraint on the non-negative matrix factor of the feature space to make full use of the sparse self-representation information. What's more, we propose a new local discriminative feature selection clustering called feature selection based dual-graph sparse non-negative matrix factorization for local discriminative clustering (DSNMF-LDC) whose clustering effects are better. We give the objective function, the iterative updating rules and the convergence proof. Our empirical study shows that DSNMF-LDC is robust and excellent in comparison to 9 feature selection algorithms and 7 clustering algorithms in clustering accuracy (ACC) and normalized mutual information (NMI). [ABSTRACT FROM AUTHOR]

With the introduction of spatial-spectral fusion and deep learning, the classification performance of hyperspectral imagery (HSI) has been promoted greatly. For some widely used datasets, the classification accuracy almost reaches 100%. However, for hyperspectral image classification, random sampling is still the most common strategy to collect the training and test samples. Because the training and test samples are randomly selected from the same image, so they have a high correlation and the classification results are overoptimistic. Besides, random sampling is not a good choice for practical applications because we cannot always collect training and test samples from the same region. Disjoint sampling selects training and testing samples from different local regions, which will provide a more objective performance evaluation for HSI classification models. In this paper, we first show the huge classification performance difference caused by different sampling strategies with a simple experiment, then we analyze the underlying reasons from the spectral information, spatial-spectral combination, sample overlapping and spatial distance, finally, a semi-supervised feature learning method is proposed for disjoint HSI classification, in which the spatial and spectral information are exploited effectively and reasonably. The experimental results based on three HSI datasets demonstrate the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

Most of the existing unsupervised feature selection methods learn the cluster structure through spectral clustering, and then use various regression models to introduce the data matrix into the indicator matrix to obtain feature selection matrix. In these methods, the clustering indicator matrix is usually continuous value, which is not the best choice for the matrix in terms of its supervising role in feature selection. Based on this, unsupervised feature selection via discrete spectral clustering and feature weights (FSDSC) is proposed in this paper. First, FSDSC integrates regression model and spectral clustering in a unified framework for feature selection, and introduces a feature weight matrix, which intuitively expresses the importance of each feature with its diagonal elements. Compared with the common feature selection matrix that requires constraints such as sparse regular items, the appearance of the feature weight matrix reduces the complexity of the model and simplifies the calculation process of feature evaluation. Secondly, for the value of the indicators matrix, the spectral clustering is improved to obtain a discrete clustering indicator matrix, which provides clearer guidance information for feature selection. Finally, in order to avoid trivial solutions, the transformation matrix is constrained by orthogonal constraint. The combination of the orthogonal regression model and spectral clustering enables the algorithm to perform feature selection and manifold information learning at the same time, thereby preserving the local geometric structure of data. Compared with other excellent unsupervised feature selection algorithms, the experimental results prove the effectiveness of the proposed algorithm. [ABSTRACT FROM AUTHOR]

Subspace representation based salient object detection has received increasing interests in recent years. However, due to the independent coding process of sparse reconstruction, the locality and the similarity among regions to be encoded are not explored. To preserve the locality and similarity of regions, a graph Laplacian regularization term is constructed as a smooth operator to alleviate the instability of the salient score in visual object. Then a new saliency map is calculated by incorporating this local graph regularizer into sparse reconstruction, which explicitly explores the local spatial structure of salient objects and thus obtains more uniform salient map. Moreover, we advance a heuristic object based dictionary from background superpixels, by which objects can be more accurately located. Experimental results on four large benchmark databases demonstrate that the proposed method performs favorably against fifteen recent state-of-the-art methods in terms of five evaluation criterions. [ABSTRACT FROM AUTHOR]

The advantage of using multifeature information for classification has been widely recognized. Representation-based methods with multifeature combination learning have only recently attracted increasing attention for hyperspectral classification. However, nonlinearity in data and the computational load of processing multifeature information and contextual information have been two thorny issues. In this paper, we present a fast joint sparse representation model with multifeature combination learning and its kernel extensions for hyperspectral imagery classification. For several complementary features (spectral, shape, and texture), the proposed model simultaneously acquires a representation vector for each type of feature and encourages the representation vectors to share a common sparsity pattern by imposing the joint sparsity ℓ r o w , 0 -norm regularization. Thus, the cross-feature information can be taken into account. For different features, different weights are assigned since they may not contribute equally to the final decision. Furthermore, kernel joint sparse representation model is presented to handle nonlinearity in the data. Kernel model projects the data into a high-dimensional space to improve the separability, achieving a better performance than the linear version. At the same time, we incorporate contextual neighborhood knowledge into the learned models. Experiments on several real hyperspectral images indicate that the proposed algorithms with much less memory requirements perform significantly faster than state-of-the-art algorithms, while exhibit highly competitive classification accuracy. [ABSTRACT FROM AUTHOR]

Feature selection algorithms eliminate irrelevant and redundant features, even the noise, while preserving the most representative features. They can reduce the dimension of the dataset, extract essential features in high dimensional data and improve learning quality. Existing feature selection algorithms are all carried out in data space. However, the information of feature space cannot be fully exploited. To compensate for this drawback, this paper proposes a novel feature selection algorithm for clustering, named self-representation based dual-graph regularized feature selection clustering (DFSC). It adopts the self-representation property that data can be represented by itself. Meanwhile, the local geometrical information of both data space and feature space are preserved simultaneously. By imposing the l 2,1 -norm constraint on the self-representation coefficients matrix in data space, DFSC can effectively select the most representative features for clustering. We give the objective function, develop iterative updating rules and provide the convergence proof. Two kinds of extensive experiments on some datasets demonstrate the effectiveness of DFSC. Extensive comparisons over several state-of-the-art feature selection algorithms illustrate that additionally considering the information of feature space based on self-representation property improves clustering quality. Meanwhile, because the additional feature selection process can select the most important features to preserve the intrinsic structure of dataset, the proposed algorithm achieves better clustering results compared with some co-clustering algorithms. [ABSTRACT FROM AUTHOR]

• This paper proposes a new feature selection method called SLMEA. • The sparse transformation and pseudo label matrix are combined to optimize. • The pseudo label matrix is used to guide the learning of low-dimensional space. • The similarity matrix is constructed adaptively. • The constraint of ℓ 2 , 1 / 2 -norm is applied to the matrix. • The SLMEA can have higher clustering accuracy and normalized mutual information. Traditional feature selection algorithms usually explore the relationship between data and cluster structure in a single space, so the internal relationship obtained is not very rich, and it is not enough to select more valuable features. To solve the above problems, in order to fully mine the intrinsic correlation information in different spaces, so that the relationship between data, features and clustering structure can be explored at the same time, this paper proposes a feature selection method based on sparse and low-dimensional representation with maximum entropy adaptive graph (SLMEA). Firstly, the SLMEA combines the sparse transform representation with pseudo-label matrix learning to optimize, and uses the pseudo-label matrix to guide the learning of sparse low-dimensional space. It can not only explore the relationship between data and pseudo-labels in data space, but also mine the association between features and pseudo-labels in feature space, so as to select more discriminative features. Secondly, based on the maximum entropy theory, the similarity matrix is constructed adaptively, so that the two different manifold structures corresponding to sparse transform representation and pseudo label matrix learning can be adaptively learned and retained in the iterative process. In addition, in order to ensure the sparsity of the transformation matrix, the constraint of ℓ 2 , 1 / 2 -norm is applied to the matrix, which can better deal with the redundant features and obtain more sparse solutions. Finally, the SLMEA uses an alternate iterative update method to optimize the objective function, and carries out extensive experiments on eight mainstream datasets. Compared with seven state-of-the-art algorithms, SLMEA can have higher clustering accuracy and normalized mutual information. [ABSTRACT FROM AUTHOR]

Abstract: Semi-supervised learning is a popular machine learning technique where only a small number of labeled examples are available and a large pool of unlabeled examples can be obtained easily. In co-training by committee, a paradigm of semi-supervised learning, it is necessary to pick out a fixed number of most confident examples according to the ranking of class probability values at each iteration. Unfortunately, the class probability values may repeat, which results in the problem that some unlabeled instances share the same probability and will be picked out randomly. This brings a negative effect on the improvement of the performance of classifiers. In this paper, we propose a simple method to deal with this problem under the intuition that different probabilities are crucial. The distance metric between unlabeled instances and labeled instances can be combined with the probabilities of class membership of committee. Two distance metrics are considered to assign each unlabeled example a unique probability value. In order to prove that our method can get higher-quality examples and reduce the introduction of noise, a data editing technique is used to compare with our method. Experimental results verify the effectiveness of our method and the data editing technique, and also confirm that the method for the first distance metric is generally better than the data editing technique. [Copyright &y& Elsevier]

In the last decade, there has been continuously increasing attention on Neural Architecture Search (NAS). The design of network architecture is aimed at automatically generating efficient neural networks in the absence of prior knowledge. Most of current oneshot NAS methods are based on weight sharing. However, there are two main problems that lead to suboptimal search results. Firstly, the large scale supernet results in a complex and difficult search for optimal subnetwork structures in the large discrete search space. Secondly, weight sharing makes the internal weight parameters of the supernet cannot always vary towards the final weights required for the optimal network structure due to the influence of different structures. These problems are responsible for the fact that the performance of the subnetworks searched by the supernet cannot represent the real performance of the subnetworks trained from scratch. To solve these problems, we propose the method based on quantum evolution and balance pool, called BQNAS, which consists of two stages. The single path supernet is trained based on weight sharing and balance pool sampling method. Then, quantum parallelism is exploited for one-hot encoding and searching for the optimal subnetwork structure based on quantum evolutionary algorithms. It can greatly improve the search efficiency of oneshot approaches in subnetwork sampling and evaluate the real performance for searching for superior subnetwork structure. Extensive experiments on two datasets show that the proposed approach outperforms the state-of-the-art ones. Specifically, BQNAS achieves a top-1 accuracy of 97.27% on CIFAR-10 and 81.36% on CIFAR-100. [ABSTRACT FROM AUTHOR]

Abstract: Eye detection plays an important role in face recognition because eyes provide distinctive facial features. However, illumination effects such as heavy shadows and drastic lighting change make it difficult to detect eyes well in facial images. In this paper, we propose a novel framework for illumination invariant eye detection under varying lighting conditions. First, we use adaptive smoothing based on the retinex theory to remove the illumination effects. Second, we perform eye candidate detection using the edge histogram descriptor (EHD) on the illumination normalized-facial images. Third, we employ the support vector machine (SVM) classification for eye verification. Finally, we determine eye positions using eye probability map (EPM). Experimental results on the CMU-PIE, Yale B, and AR face databases demonstrate that the proposed method achieves high detection accuracy and fast computation speed in eye detection. [Copyright &y& Elsevier]

Weakly supervised object localization (WSOL) is a challenging task that aims to localize objects in images using only image-level labels. Despite the widespread use of WSOL methods based on class activation mapping (CAM), such methods do not consider that the network may overly focus on local regions of the most interesting objects during localization, thus neglecting the overall information. To address this issue, we introduces an additional attention branch for convolutional neural networks (CNNs) that utilizes the attention mechanism of multi-layer perceptron (MLP) to enhance the network's learning of global information and supervise the feature learning of CNNs online through knowledge distillation, thereby improving the localization accuracy of WSOL. Specifically, we designs a new loss function using the generated features to combine with contrastive learning, effectively dividing the foreground and background of the image to provide more accurate pseudo-labels for subsequent classification and localization tasks. In the experiments, we tested our method on the CUB-200-2011 dataset and compared it with existing methods. The experimental results show that our method achieves good performance in WSOL tasks. [ABSTRACT FROM AUTHOR]

Pedestrian detection has emerged as a fundamental technology for autonomous cars, robotics, pedestrian search, and other applications. Although a lot of progress has been made in pedestrian detection, it is still a challenging problem in wide field-of-view surveillance videos due to the random distribution and dynamic characteristics of pedestrians. Especially in large-scale high-resolution images, there are many pedestrians and vast scale variance, so it is difficult to accurately detect all pedestrians. In order to solve this problem, sliding window is used to crop all original images to obtain pre-detection results firstly in this paper. Then, the original images are cropped again with the object as the center utilizing the label files shared in the same scene to get multi-scale images. Finally, a region NMS algorithm, a fusion strategy about the results of small images mapped into large images, is proposed to remove the redundant shredded boxes caused by cropping image. We verify the effectiveness of the proposed method with Faster R-CNN, Cascade R-CNN, IterDet and Scale-aware Fast R-CNN models on PANDA dataset. [ABSTRACT FROM AUTHOR]

Abstract: Classification is very difficult in high-dimensional spaces because learning methods suffer from the curse of dimensionality. In order to efficiently classify the high-dimensional data, a Supervised Immune Clonal Evolutionary Classification Algorithm (SICECA) is proposed in this paper. First, the automatic nonparametric Uncorrelated Discriminant Analysis (UDA) is adopted for Dimensionality Reduction (DR), which combines rank-preserving dimensionality reduction with constraint discriminant analysis so as to realize the extracted features statistically uncorrelated. Then, an Immune Clonal Evolutionary Algorithm (ICEA) based on clonal selection principle in immunology is proposed to act as classifier. In the experiments, first of all, 11 UCI data sets, four texture images and three Synthetic Aperture Radar (SAR) images are used to test the performance of SICECA. SICECA is also compared with three existing algorithms in terms of classification accuracy and running time. In addition, SICECA is applied to a real world application, namely, face recognition, with a good performance obtained. [Copyright &y& Elsevier]

Abstract: The goal of segmentation is to partition an image into disjoint regions. In this paper, the segmentation problem based on partition clustering is viewed as a combinatorial optimization problem. A new algorithm called a quantum evolutionary clustering algorithm based on watershed (QWC) is proposed. In the new algorithm, the original image is first partitioned into small pieces by watershed algorithm, and the quantum-inspired evolutionary algorithm is used to search the optimal clustering center, and finally obtain the segmentation result. Experimental results show that the proposed method is effective for texture image and SAR image segmentation, compared with QICW, the genetic clustering algorithm based on watershed (W-GAC) and K-means algorithm based on watershed (W-KM). [Copyright &y& Elsevier]

Abstract: Based on the previous work on ridgelet neural network, which employs the ridgelet function as the activation function in a feedforward neural network, in this paper we proposed a single-hidden-layer regularization ridgelet network (SLRRN). An extra regular item indicating the prior knowledge of the problem to be solved is added in the cost functional to obtain better generalization performance, and a simple and efficient method named cost functional minimized extreme and incremental learning (CFM-EIL) algorithm is proposed. In CFM-EIL based SLRRN (CFM-EIL-SLRRN), the ridgelet hidden neurons together with their parameters are tuned incrementally and analytically; thus it can significantly reduce the computational complexity of gradient based or other iterative algorithms. Some simulation experiments about time-series forecasting are taken, and several commonly used regression ways are considered under the same condition to give a comparison result. The results show the superiority of the proposed CFM-EIL-SLRRN to its counterparts in forecasting. [Copyright &y& Elsevier]

Abstract: The multiscale properties of the reception field of the human visual cortex have illuminated the research of wavelet neural network (WNN). Findings in neurophysiology indicate that in the human visual system there are specialized areas in the visual cortex that respond for particular orientations. In other words, the reception field of visual cortex has the multiresolution properties in direction, as well as in localization and scale. Enlightened by these facts, a three layer feed-forward neural network (FNN) is presented by employing ridgelet as the activation function in the hidden layer. To get rapid learning when dealing with high dimensional samples, we proposed an efficient linear learning algorithm inspired by traditional kernel smoothing method, which has low computation complexity proportional to the number and dimension of samples. At the cost of a little degradation in accuracy, the network can achieve rapid learning. Some simulation experiments about function approximation are taken, and several commonly used regression ways are considered under the same conditions to give a comparison result. The results show that the proposed linear ridgelet network can overcome the curse of dimensionality in the training of FNNs and exhibit better performance in high dimension than its counterparts, especially when some spatial inhomogeneities exist in the function. [Copyright &y& Elsevier]

Abstract: In this paper, a new kind of neural network is proposed by combining ridgelet with feedforward neural network (FNN). The network adopts ridgelet as the activation function in the hidden layer, and an incremental constructive method is employed to determine the structure of the network. Since ridgelets are efficient in describing linear, curvilinear, and hyperplane like structures in high dimensions, accordingly the network can approximate quite a wide range of multivariate functions in a more stable and efficient way, especially those with certain kinds of spatial inhomogeneities. Moreover, the incremental extreme learning machine makes adding the hidden nodes one by one possible, and it only needs to adjust the output weights linking the hidden layer and the output layer when more hidden neurons are added. By defining the cost function as the difference between the previously approximated function and the currently approximating one, a genetic algorithm is used to determine the optimal directions of ridgelet neurons. The construction and learning of the network are presented in detail. The superiority of the proposed model is demonstrated by simulation experiments in function learning and image compression. [Copyright &y& Elsevier]

Abstract: A ridgelet kernel regression method is presented in this paper to approximate multi-dimensional functions, especially those with certain kinds of spatial inhomogeneities. This method is based on ridgelet theory, kernel and regularization techniques from which we can deduce a regularized kernel regression form. By representing this form with quadratic programming and taking the obtained solution to define a fitness function, we use particle swarm optimization to optimize the directions of ridgelets. The properties of ridgelet can guarantee the stability of this method in approximating multi-dimensional functions, as well as its superiority for functions with linear singularities. Additionally, the regularized technique employed in this model leads to smaller generalization error. Experiments in the tasks of regression and classification show its effectiveness. [Copyright &y& Elsevier]

Knowledge distillation (KD) builds a lightweight Student Model (SM) and trains it to approximate a large Teacher Model (TM) by exploring knowledge learned by the TM, which shows effectiveness to train lightweight CNN models. However, training a small SM to achieve better performance remains a challenging problem. Recent researches on human learning behaviors show that both the knowledge from teachers and the knowledge learning processes of teachers are significant for students. Inspired by this characteristic, in this paper, we propose a new perspective, called Progressive Mimic Learning (PML), to train lightweight CNN models by mimicking the learning trajectory of the TM. In order to obtain a more powerful SM, the useful hints in the learning process of the TM are explored. To start with, the TM learning process is divided into multiple stages, and the last state of the TM in each stage is recorded as a landmark. The learning trajectory of the TM is composed of these landmarks. Then, a landmark loss is defined to constrain the SM to progressively mimic the learning process of the TM, by employing landmarks in the learning trajectory as a training hint of the SM. Several experiments are conducted on four benchmark data sets, CIFAR-10, CIFAR-100, Fashion-MNIST, and ImageNet-10, to investigate the performance of the PML. The results show that the PML can make SMs generate more accurate predictions than SMs trained by its counterparts. [ABSTRACT FROM AUTHOR]

Spectral–spatial classification methods can improve the classification accuracy of hyperspectral imagery (HSI) dramatically. However, this improvement is caused in part by the overlap between training set and test set. In this paper, a novel non-overlapping spectral-spatial classification framework based on set-to-sets distance is proposed. First, each class is sampled using a controlled random sampling method, which is regarded as a training set. The image is segmented into many superpixels and each superpixel is taken as a test set. In order to ensure that the test set and the training set are not overlapped with each other, the training pixels contained in each test superpixel are deleted. Then, the training set of each class is compressed into a more compact set to reduce the computational complexity and kernel trick is used to make samples be approximately linearly separable. Finally, each test set is modeled as a convex hull, this hull is represented collaboratively with all training sets. With the resolved representation coefficients, the distance between the test set and each training set can be calculated for classification. Experimental results based on three real HSI data sets demonstrate the superiority of the proposed method to state-of-the-art algorithms under the non-overlapping sampling strategy. [ABSTRACT FROM AUTHOR]

Convolutional neural network (CNN) has proven effective at solving difficult image classification problems, but it can be challenging to design its architecture. We try to simplify the search process of the optimal architecture and minimize human participation, so quantum behaved Particle Swarm Optimization with binary encoding (BQPSO) is employed after analyzing the limitations of traditional Particle Swarm Optimization (PSO). To do this, we propose a novel and robust binary encoding strategy which does not require users with domain knowledge on CNN. Then a new quantum behaved evolving strategy is proposed to ensure the effectiveness of evolved CNN architectures. Finally, the performance of our algorithm is measured by the classification accuracy for several benchmark datasets commonly used in deep learning. The experimental results prove that our proposed method can achieve better performance and robustness than the traditional method. This is the first completely automatic algorithm in the area of using quantum behaved PSO to evolve CNN architectures. [ABSTRACT FROM AUTHOR]

The high dimensionality of hyperspectral imagery often introduces challenge for the conventional data analysis techniques. In order to improve the classification performance of hyperspectral imagery, metric learning is often introduced to assign small distances between samples from the same class and large distances from different class. However, most of the traditional metric learning methods only adopt linear transformations, which cannot capture the complex nonlinear relationships between high dimensional samples. Inspired by the successful application of deep learning for the classification of hyperspectral imagery. In this paper, the deep neural network is introduced to learn the discriminating metric for the classification of hyperspectral images. In order to improve the reliability of classification results, the spectral and spatial information are combined with weighted classification probabilities. Experimental results demonstrate that the proposed method achieves satisfactory classification performance when compared with other metric learning methods or deep models. [ABSTRACT FROM AUTHOR]

• A precise crop classification method of UAV hyperspectral imagery is proposed. • KTSSCC can retain the spectral and spatial information while improving the linear separability. • The KTSSOMP algorithm adaptively learns the sparse coding coefficients of KTSSCC. • There are very few parameters in our proposed KTSSCC model. Precise crop classification plays a significant role in the agriculture field. An appropriate data source for precise crop classification is high spatial resolutions hyperspectral imagery (H2 imagery) acquired by unmanned aerial vehicle (UAV). However, for imagery with many different classes of crops, crop classification of UAV H2 imagery is a huge challenge. The significant spectral diversity, spatial heterogeneity and nonlinear data structure of UAV H2 imagery results in poor spectral discriminability. To improve the discriminability, a kernel tensor slice sparse coding-based classifier (KTSSCC) is proposed for precise crop classification of UAV H2 imagery in this research. The kernel tensor representation mechanism in KTSSCC can reduce the nonlinear separation while well preserving the spectral characteristics and spatial constraints of land-covers, and thus the discriminability is greatly improved. Furthermore, this paper puts forward the kernel tensor slice sparse orthogonal matching pursuit (KTSSOMP) algorithm to optimize kernel tensor slice sparse coding in the spectral space, which greatly reduces the computation cost. Moreover, there are very few parameters to be tuned in our proposed model. We assess the performance of KTSSCC on two real UAV hyperspectral imagery datasets, and find that, based on visual and quantitative results, it provides satisfactory crop classification results and outperforms the state-of-the-art approaches. [ABSTRACT FROM AUTHOR]

The current self-examples learning is a promising direction for image super-resolution (SR) task, but still has several main drawbacks: (i) less ability of adapting the neighboring and non-local information for self-regression function learning; (ii) less priors from the training data generated by recursively down-sampling the low resolution (LR) input. In this work, we propose an external learning assisted self-examples learning SR (Exter-SESR) framework to alleviate these issues, which is conducted in a two-stage procedure. The first part is a hybrid neural network (HNN) that takes large LR image patches as input and extracts compact features from external dataset. The learned features can offer better neighboring and non-local priors. Meanwhile, a part of HNN is able to estimate an initial high resolution (HR) image to address the second issue, since the new training data from the initial HR images does not only preserve the original prior, but also involve the extra ones from the external data. The second part is to further refine the SR output using a SESR based model. In addition, we analyze the effects of different self-examples models on the SR performance and find that Gaussian process regression (GPR) achieves superior performance. Experimental results on the benchmark show that our proposed method outperforms the existing SESR methods by a large margin in terms of both quantitative and qualitative measurements. [ABSTRACT FROM AUTHOR]

Matrix factorization (MF) is an increasingly important approach in the field of missing value prediction because recommender systems are rapidly becoming ubiquitous. MF-based collaborative filtering (CF) seeks to improve recommender performance by combining user–item matrix with MF. However, most MF-based approaches available at present could not obtain high prediction accuracy because of the sparse availability of user–item matrices in CF models. The present paper proposes a framework that involves two efficient MF, dynamic single-element-based CF-integrating manifold regularization (DSMMF) and dynamic single-element-based Tikhonov graph regularization non-negative MF (DSTNMF). The aim of this framework is to better use the intrinsic structure of user–item rating matrix and user/item content information, overcome the dimensionality curse and ill-posed problem of weighted graph NMF, and evade the frequent manipulations of indicator matrices that lack practicability. We validate the effectiveness of our proposed algorithms with respect to recommender performance by four indices on three datasets. We demonstrate that our proposed approaches lead to considerable improvement compared with several other state-of-the-art approaches. [ABSTRACT FROM AUTHOR]

Sparse coding based Single Image Super-Resolution (SISR) technology has proved to be effective in generating High-Resolution (HR) image from a single Low-Resolution (LR) image. However, unsuitable codebook will bring unexpected structural details in the resultant HR images. In this paper, in order to find more reliable image features for formulating codebook, we propose new Curvelet Support Value Filters (CSVFs) for multiscale structural features extraction. By defining a local multi-scale and multi-directional Curvelet function to approximate images and casting a structural risk minimization constraint in this approximation, we can derive a set of Curvelet support values to reveal the most salient local scale-location-direction information of images. Then these features are used to construct more reliable codebook, and sparse coding based recovery is performed at multiple scale and directions via two coupled example datasets, to recover HR images. Some experiments are taken on realizing a 3X amplification of natural images, and the recovered results suggest its efficiency and superiority to its counterparts. [ABSTRACT FROM AUTHOR]

Many researchers have introduced tag information to recommender systems to improve the performance of traditional recommendation techniques. However, user-defined tags will usually suffer from many problems, such as sparsity, redundancy, and ambiguity. To address these problems, we propose a new recommendation algorithm based on deep neural networks. In the proposed algorithm, users׳ profiles are initially represented by tags and then a deep neural network model is used to extract the in-depth features from tag space layer by layer. In this way, representations of the raw data will become more abstract and advanced, and therefore the unique structure of tag space will be revealed automatically. Based on those extracted abstract features, users׳ profiles are updated and used for making recommendations. The experimental results demonstrate the usefulness of the proposed algorithm and show its superior performance over the clustering based recommendation algorithms. In addition, the impact of network depth on the algorithm performance is also investigated. [ABSTRACT FROM AUTHOR]

In this paper, a novel method quantum clustering using kernel entropy component analysis (KECA-QC) is proposed. This method has two phases: preprocessing and clustering stages. The main idea of preprocessing is to map the original data to a high-dimensional feature space, and to select the useful components using Renyi entropy as our similarity metric. After data preprocessing, different clusters will be distributed more or less in different places, and for high-dimensional datasets, it can achieve the purpose of dimensionality reduction at the same time. In the second phase, quantum clustering method is used, which can find clusters of any shape without knowing the number of clusters. Based on the traditional quantum clustering, we develop a new method estimating the wave function from distributions of K- nearest neighbors statistics, which can further reduce the running time and improve the calculation efficiency. In order to evaluate the effectiveness of this method, we compare the proposed method with k -means clustering (KM), the classical spectral clustering algorithm called Ng–Jordan–Weiss (NJW), the traditional QC, and kernel entropy component analysis spectral clustering algorithm (KECA-KM). The experimental results demonstrate that the proposed algorithm outperforms the compared algorithms on synthesized datasets and UCI datasets. [ABSTRACT FROM AUTHOR]

In this paper, an Analog–Digital Mixing Network (ADMN) is advanced for simultaneously collecting data and classifying the Power Quality (PQ) events. Based on recently developed Compressed Sampling (CS) theory, power signals are sampled via a new robust and semi-supervised compressive sampling scheme, and then the recorded data are directly used as features for the subsequent classification. Moreover, an Online Sequential Learning Algorithm (OSLA) is proposed to learn the training data one-by-one or chunk by chunk, and discard them as long as the training procedure is completed to keep the memory bounded in online learning. Consequently, ADMN can collect data streams and classify them sequentially, which provides a promising way to deal with the “big data”. Some experiments are taken on the classification of real PQ events, and the experimental results show the efficiency and superiority of our proposed method to its counterparts. [ABSTRACT FROM AUTHOR]

Making a close inspection of the recent progress made in computer vision we will find that more and more advancement can be attributed to the introduction of bio-inspired algorithms, which is a flourishing area of computing. Biological patterns can be generated from salient attention principle of the nervous system, which pops out the most important information from large backgrounds. New evidences of the underlying neuronal mechanisms indicate that biological motion phenomena happen on higher cognition level instead of low-level pixels in sensing environment, especially in a complex scene. Inspired by it, this paper proposes a novel saliency detection method based on a directed graph model and multi-scale Bayesian inference. We first create a directed graph with superpixels as its nodes and introduce a baseline node whose saliency is considered to be zero. The saliency of each node is defined as the shortest distance from the baseline node to it and Dijkstra׳s algorithm is adjusted to solve this optimization problem with great efficiency. Furthermore, we extend this model to a multi-scale version to cope with salient regions of different sizes and a Bayesian inference strategy via 3D color histogram is developed to achieve pixel-level saliency. Experimental results on some benchmark dataset demonstrate the superiority of our method with respect to 18 state-of-the-art saliency detection methods and our method achieves the highest recall in MSRA-1000. Additional experiments on ship detection of SAR images show that the proposed method can overcome the shortcomings of traditional CFAR detector and has much fewer false alarms in cluttered background. [ABSTRACT FROM AUTHOR]

Multiobjective evolutionary algorithm based on decomposition (MOEA/D) has attracted a great deal of attention and has obtained enormous success in the field of evolutionary multiobjective optimization. It converts a multiobjective optimization problem (MOP) into a set of scalar optimization subproblems and then uses the evolutionary algorithm (EA) to optimize these subproblems simultaneously. However, there is a great deal of randomness in MOEA/D. Researchers in the field of evolutionary algorithm, reinforcement learning and neural network have reported that the simultaneous consideration of randomness and opposition has an advantage over pure randomness. A new scheme, called opposition-based learning (OBL), has been proposed in the machine learning field. In this paper, OBL has been integrated into the framework of MOEA/D to accelerate its convergence speed. Hence, our proposed approach is called opposition-based learning MOEA/D (MOEA/D-OBL). Compared with MOEA/D, MOEA/D-OBL uses an opposition-based initial population and opposition-based learning strategy to generate offspring during the evolutionary process. It is compared with its parent algorithm MOEA/D on four representative kinds of MOPs and many-objective optimization problems. Experimental results indicate that MOEA/D-OBL outperforms or performs similar to MOEA/D. Moreover, the parameter sensitivity of generalization opposite point and the probable to use OBL is experimentally investigated. [ABSTRACT FROM AUTHOR]

The traditional reproduction operators, which are originally designed for single-objective optimization, are directly adopted in most state-of-the-art multi-objective evolutionary algorithms (MOEAs). However, these reproduction operators might not be suitable for multiobjective optimization problems (MOPs) due to the regularity property of continuous MOP, and that is to say its Pareto optimal set in the decision space is generally piecewise continuous manifold rather than a set of independent points. Few researches have used this regularity property of continuous MOP to help design their algorithms. In this paper, based on the regularity property, a Baldwinian learning operator is incorporated into the framework of MOEA/D (multi-objective evolutionary algorithm based on decomposition) and thereby we propose MOEA/D/BL. The Baldwinian learning operator obtains the evolving information based on the learned distribution model of a current population. It constructs a candidate descent direction based on the learned distribution model and the evolving history of the parent individuals. Experimental results on twenty-three popular test problems show that the proposed algorithm performs better than or as well as four other compared algorithms. It also experimentally demonstrates that the proposed Baldwinian learning operator can accelerate the convergence of solutions. [ABSTRACT FROM AUTHOR]

Abstract: In this paper, a Sparse Ridgelet Kernel Regressor (SRKR) is constructed by combing the ridgelet theory, the sparse theory with kernel trick. By using the dimensionality non-separable ridgelet kernels, SRKR is capable of processing the high-dimensional data more efficiently. Considering the preferable future of sequential learning over batch learning in problems where data arrive constantly and where batch learning is expensive, we exploit the new kernel method in an online setting using the sequential extreme learning scheme. The online learning algorithm of the examples, named Online Sequential Extreme Learning Algorithm (OS-ELA) is employed to rapidly produce a sequence of estimations. OS-ELA learns the training data one-by-one or chunk by chunk (with fixed or varying size), and discards them as long as the training procedure for those data is completed to keep the memory bounded in online learning. Evolution scheme is also incorporated to obtain a ‘good’ sparse regressor. Experiments are taken on some nonlinear time-series prediction problems, in which the examples are available one by one. Some comparisons are made and the experimental results show its efficiency and superiority to its counterparts. [Copyright &y& Elsevier]

Abstract: Graph-based semi-supervised learning has been widely researched in recent years. A novel Low-Rank Representation with Local Constraint (LRRLC) approach for graph construction is proposed in this paper. The LRRLC is derived from the original Low-Rank Representation (LRR) algorithm by incorporating the local information of data. Rank constraint has the capacity to capture the global structure of data. Therefore, LRRLC is able to capture both the global structure by LRR and the local structure by the locally constrained regularization term simultaneously. The regularization term is induced by the locality assumption that similar samples have large similarity coefficients. The measurement of similarity among all samples is obtained by LRR in this paper. Considering the non-negativity restriction of the coefficients in physical interpretation, the regularization term can be written as a weighted -norm. Then a semi-supervised learning framework based on local and global consistency is used for the classification task. Experimental results show that the LRRLC algorithm provides better representation of data structure and achieves higher classification accuracy in comparison with the state-of-the-art graphs on real face and digit databases. [Copyright &y& Elsevier]