Advances in Brain Inspired Cognitive Systems

Breathtaking improvement has been seen with the recently proposed deep Generative Adversarial Network (GAN). Purposes of most existing GAN-based models majorly concentrate on generating realistic and vivid patterns by a pattern generator with the aid of the binary discriminator. However, few study were related to the promotion of classification performance with merits of those generated ones. In this paper, a novel and generalized classification framework called Style Neutralization Generative Adversarial Classifier (SN-GAC), based on the GAN framework, is introduced to enhance the classification accuracy by neutralizing possible inconsistent style information existing in the original data. In the proposed model, the generator of SN-GAC is trained by mapping the original patterns with certain styles (source) to their style-neutralized or standard counterparts (standard-target), capable of generating the targeted style-neutralized one (generated-target). On the other hand, pairs of both standard (source + standard-target) and generated (source + generated-target) patterns are fed into the discriminator, optimized by not only distinguishing between real and fake, but also classifying the input pairs with correct class label assignment. Empirical experiments fully demonstrate the effectiveness of the proposed SN-GAC framework by achieving so-far the highest accuracy on two benchmark classification databases including the face and the Chinese handwriting character, outperforming several relevant state-of-the-art baseline approaches.

The universe approximate theorem states that a shallow neural network (one hidden layer) can represent any non-linear function. In this paper, we aim at examining how good a shallow neural network is for solving non-linear decision making problems. We proposed a performance driven incremental approach to searching the best shallow neural network for decision making, given a data set. The experimental results on the two benchmark data sets, Breast Cancer in Wisconsin and SMS Spams, demonstrate the correction of universe approximate theorem, and show that the number of hidden neurons, taking about the half of input number, is good enough to represent the function from data. It is shown that the performance driven BP learning is faster than the error-driven BP learning, and that the performance of the SNN obtained by the former is not worse than that of the SNN obtained by the latter. This indicates that when learning a neural network with the BP algorithm, the performance reaches a certain value quickly, but the error may still keep reducing. The performance of the SNNs for the two databases is comparable to or better than that of the optimal linguistic attribute hierarchy, obtained by a genetic algorithm in wrapper or in terms of semantics manually, which is much time-consuming.

Many studies have shown that there is a decline in seminal quality during the past two decades. Seminal quality may be affected by environmental factors and health status, as well as life habits. Artificial intelligence (AI) technology has been recently applied to recognize this effect. However, conventional AI algorithms are not prepared to cope with the class-imbalanced fertility dataset. To this end, a back-propagation neural network (BPNN) is used to predict the seminal profile of an individual from the dataset. A neural-genetic algorithm (N-GA) is employed to select optimal feature subsets and optimize the parameters of the used neural network. Results indicate that the proposed method outperforms other AI methods on seminal quality prediction in terms of precision and accuracy.

In recent years, with the rapidly development of Internet and pharmaceutical market, online medical platform has become a major place for online medical trading. Recommendation systems have been widely deployed in commercial platform to improve user experience and sales. Motivated by this, we propose two hybrid recommendation algorithms, CB-CF hybrid algorithm and CNN-based CF algorithm, for B2B medical platform to provide accurate recommendations. We also give a brief introduction of two well-known recommendation algorithms, content-based algorithm and model-based CF algorithm. Then we investigate the performance of recommendation algorithms on Apache Spark and Tensorflow with real-world data collected from a china B2B online medical platform. Experimental results show that the hybrid recommendation algorithm performs better than other algorithms.

Through introducing the psychology model of reading cognitive, this paper uses the LSTM neural network and the CRF model respectively to simulate the language cognition process and the eye-movement control process in reading, in order to overcome the defect that the traditional CRF prediction model only considers the context information of the label sequence but can not take into account the context information of the text sequence. First, the psychological process of reading cognition is introduced and the prediction model of saccade target based on LSTM-CRF for Chinese reading is proposed. Then, the experimental data, experimental environment, feature templates and parameter settings needed for model training are introduced. Finally, the conclusion is drawn through experimental comparison: (1) The F1 score of prediction model in saccade labeling based on LSTM-CRF is superior to the traditional CRF prediction model; (2) The predictability of the language itself is an important feature of the saccade target prediction model; (3) The best saccade length for Chinese readers is about 2.5 Chinese characters.

Vehicle re-identification has gradually gained attention and widespread applications. However, most of the existing methods learn the discriminative features for identities by single feature channel only. It is worth noting that visual cognition of human eyes is a multi-channel system. Therefore, integrating the multi-view information is a nature way to boost computer vision tasks in challenging scenarios. In this paper, we propose to mine multi-view deep features via correlative sparse ranking for vehicle re-identification. Specifically, first, we employ ResNet-50 and GoogleNet as two baseline networks to generate the attributes (vehicle color and type) aggregated features. Then we explore the feature correlation via enforcing the correlation term into the multi-view sparse coding framework. The original rankings are obtained by the reconstruction coefficients between probe and gallery. Finally, we utilize a re-ranking technique to further boost the performance. Experimental results on public benchmark VeRi-776 dataset demonstrate that our approach outperforms state-of-art approaches.

The synapse, which is the carrier of neurotransmitter molecules to transmit and store information, is believed to be the key to the reconstruction of the neural circuit. To date, electron microscope (EM) is considered as one of the most important tools for observing and analyzing synaptic structures because they can clearly observe the internal structure of cells. Consequently, many meaningful researches are focused on how to detect and segment the synapses from EM images. In this paper, we propose a novel and effective method to automatically detect and segment the synaptic clefts by using Mask R-CNN. On this base, we utilize the context cues in adjacent sections to eliminate the misleading results. We apply the method to the CREMI challenge and the results demonstrate that our method is effective in segmenting the synaptic clefts of the drosophila. Specifically, we rank first in sample B+ dataset, and the CREMI score is 86.50 which outperforms most of state-of-the-art methods by a large margin.

In this paper, we introduce an efficient framework to subtract the background from both visible and thermal imagery for pedestrians’ detection in the urban scene. We use a deep neural network (DNN) to train the background subtraction model. For the training of the DNN, we first generate an initial background map and then employ randomly 5% video frames, background map, and manually segmented ground truth. Then we apply a cognition-based post-processing to further smooth the foreground detection result. We evaluate our method against our previous work and 11 recently widely cited method on three challenge video series selected from a publicly available color-thermal benchmark dataset OCTBVS. Promising results have been shown that the proposed DNN-based approach can successfully detect the pedestrians with good shape in most scenes regardless of illuminate changes and occlusion problem.

Image super-resolution is an important research field in image analysis. The techniques of image super-resolution has been widely used in many computer vision applications. In recent years, the success of deep learning methods in image super-resolution have attracted more and more researchers. This paper gives a brief review of recent deep learning based methods for single image super-resolution (SISR), in terms of network type, network structure, and training methods. The advantages and disadvantages of these methods are analyzed as well.

Annotation data is the “fuel” of vision cognitive system but hard to obtain. We focus on finding a feasible way to generate high-quality image data. The 3D models can produce rich annotated 2D images, and the generative adversarial nets can create various pictures. We proposed the background augmentation generative adversarial nets to build a bridge between GAN and 3D models for data augmentation. As a result, we use BAGAN and 3D models to generate images which can help deep convolutional classifier improve accuracy score to 93.12% on real data test sets.

Image super-resolution is a process of obtaining one or more high-resolution image from single or multiple samples of low-resolution images. Due to its wide applications, a number of different techniques have been developed recently, including interpolation-based, reconstruction-based and learning-based. The learning-based methods have recently attracted increasing great attention due to their capability in predicting the high-frequency details lost in low resolution image. This survey mainly provides an overview on most of published work for single image reconstruction using Convolutional Neural Network. Furthermore, common issues in super-resolution algorithms, such as imaging models, improvement factor and assessment criteria are also discussed.

Object transfiguration aims to translate objects in image from a kind to another, which is a subtask of image translation. Recently, researchers have proposed many effective approaches for object transfiguration. However, most of them ignore the difference between target objects and background, which would make background deformation, discolor and other problems. We propose a novel attention-based model for unsupervised object transfiguration called Deep Attention Units Generative Adversarial Network (DAU-GAN). We utilize spatial consistencies of objects and background to enable model to preserve background of image. Such an attention-based design enables DAU-GAN to enhance the expression of meaningful features and let the model able to distinguish specific objects and background in images. Experimental results demonstrate that our approach improves the performance of object transfiguration as well as effectively preserves background.

Hyperspectral image classification has been widely used in a variety of applications such as land cover analysis, mining, change detection and disaster evaluation. As one of the most-widely used classifiers, the Multilayer Perception (MLP) has shown impressive classification performance. However, the MLP is very sensitive to the setting of the training parameters such as weights and biases. The traditional parameter training methods, such as, error back propagation algorithm (BP), are easily trapped into local optima and suffer premature convergence. To address these problems, this paper introduces a modified gravitational search algorithm (MGSA) by employing a multi-population strategy to let four sub-populations explore the different areas in search space and a Gaussian mutation operator to mutate the global best individual when swarm stagnate. After that, MGSA is used to optimize the weights and biases of MLP. The experimental results on a public dataset have validated the higher classification accuracy of the proposed method.

Post-earthquake high resolution (HR) remote sensing image classification is crucial for disaster assessment and emergency rescue. 3-D convolutional neural networks (3-D CNNs) exhibit promising performance in remote sensing image classification. However, 3-D CNNs lack the theoretical underpinnings to perform multiresolution approximation for filter learning in view of the scale variance of natural objects. Gabor filtering can effectively extract multiresolution spatial information including edges and textures, which have a potential to reinforce the robustness of learned features in 3-D CNNs against the orientation and scale changes. In this paper, we propose a combined 3-D convolutional neural network and Gabor filters (GNN) method for post-earthquake HR image classification. Instead of choosing a single scale, GNN extends the spatial information to several scales by Gabor filters to take advantage of correlations among multiple scales for damage mapping. The experimental results show that GNN can reflect the multiscale information of complex scenes, obtain good classification results for mapping post-earthquake damage using HR remote sensing images.

In recent years motor imagery-based brain–computer interface (MI-BCI) is widely used in the rehabilitation of stroke patients and received certain therapeutic effect. The existing imagery mode of brain-computer interface focuses more on the aspect of pure motor imagery and less on the experimental ways of combining other motion and imagination. In this paper, aiming at studying the role of attention in the context of classifying covert and overt motor activities, we design different experiments to explore it in different modes. In our experiments, covert activities are only motor imagery. Overt motor activities are divided into two types—attention to the screen and attention to intended hand. The classification accuracy of six subjects in three modes are compared and analyzed. The average accuracy of overt motor activities with attention to intended hand is the highest, which are respectively 3% and 5% higher than those of covert activities and overt motor activities with attention to screen. At the same time, overt motor activities with attention to intended hand induce more active brain areas according to the spatial pattern of the corresponding EEG data.

Image captioning, which aims to automatically generate sentences for images, has been exploited in many works. The attention-based methods have achieved impressive performance due to its superior ability of adapting the image’s feature to the context dynamically. Since the recurrent neural network has difficulties in remembering the information too far in the past, we argue that the attention model may not be adequately supervised by the guidance from the previous information at a distance. In this paper, we propose a memory-enhanced attention model for image captioning, aiming to improve the attention mechanism with previous learned knowledge. Specifically, we store the visual and semantic knowledge which has been exploited in the past into memories, and generate a global visual or semantic feature to improve the attention model. We verify the effectiveness of the proposed model on two prevalent benchmark datasets MS COCO and Flickr30k. The comparison with the state-of-the-art models demonstrates the superiority of the proposed model.

By simulating how our human brain solves complex and conceptual problems, cognitive systems have been successfully applied in a wide range of applications. In this paper, a cognitive modelling based inversion method, the direction guided differential evolution with cooperative coevolutionary mutation operator (DG-DECCM) algorithm, is proposed to trace the ray path of the seismic waves. The DE algorithm seeks optimal solutions by simulating the natural species evolution processes and makes the individuals become optimal. Classical ray tracing methods were time consuming and inefficiency. The proposed algorithm is suitable for the high and super high dimensional separable model space. It treats the emergent angles of the reflection points as genes of an individual. We introduce a sign function to guide the direction of the mutation and propose two kinds of stopping criteria for effective iteration to speed up the computation. For the complex velocity model, the local optimization methods based on gradient are time consuming to converge or may converge to local minimum but not the optimal value. The proposed global DE algorithm, however, will obtain a global optimum solution more efficiently and has higher convergence rate.

Soldier patrolling is a risky task at the cross borders which leads to loss of life. To overcome such risks, many researchers working on reduction of human effort using Cognitive Science through Brain Computer Interface (BCI) application. Human brain is a complex organ of body and researchers aim to build a direct communication of human brain with computer system including the Artificial Intelligence (AI) and Computational Intelligence (CI). In order to achieve such objectives, a proper brain signal capturing mechanism to be used. The appropriate signals are captured using Electroencephalogram (EEG) cap which is used to record electrical activity of brain and classified to filter P300 brain wave which is an Event Related Potential (ERP) to detect abnormal events like crawling under the Line of Control (LoC) or any illegal cross border movements of goods, drugs supply, arms supply and cargos. Brain signal is contaminated with artifacts and noises. Further work is carried on improving the Signal to Noise Ratio (SNR) quality by using appropriate filtration algorithm. The proposed filter is to use sliding Hierarchal Discriminant Classification Algorithm (sHDCA) for P300 signal to detect and classify between the target and non target component based on a multi Rapid Serial Visual Presentation (RSVP) using real time video frames from the region. As a result, it reduces the false alarm and creating the threat signature library from the filtered and classified brain signals for Comprehensive Integrated Border Management System (CIBMS).

In motor imagery-based Brain Computer interfaces (BCIs), the classification accuracy of using the Common Spatial Pattern (CSP) algorithm to deal with the electroencephalogram (EEG) is closely related to the frequency range selected. Due to individual differences, the frequency range selected that reaches the best performance is different, which limits the generality and the actual use of the algorithm. To solve this problem, this paper proposes a motor imagery recognition method based on Filter Bank Common Spatial Pattern (FBCSP) and Principal Components Analysis (PCA), which is called FBCSP+PCA. The feasibility of the FBCSP+CSP is preliminary verified using the 2008 BCI competition data and further verified using data collected by our laboratory with wireless dry electrode device. The average classification accuracy of the data collected by our laboratory reaches 75.7% in the absence of individual band selection. That is also to say that the proposed method has good generality and and practical value because it can obtain high performance without the need of giving each individual a specific optimum frequency band.

This paper focuses on the online implementation of a hybrid brain computer interface (BCI) involving electroculogram (EOG) and electroencephalogram (EEG) of motor imagery (MI). The hybrid BCI system comprises of modules of eye-blinking detection, ICA spatial filter, zero-training classifier and cursor movement controlling. Eye-blinking information contained in EOG signal was achieved for locating EEG segments related to motor imageries. Then, independent component analysis (ICA) was applied to the filtered EEG data to yield the motor-related potentials, whose features were fed into a zero-training classifier. Finally, the classification results regarding the types of moving imagination were transferred into commands to control the cursor moving along a predesigned path shown on the computer screen. Four subjects attended the online BCI tests, the average moving accuracy reached 84.56% for all tests, and the response time was about 4.13 trials/min. The experimental results demonstrate that the hybrid MIBCI system in this study is feasible for the real-time control of peripheral devices.

Due to the role of the brain’s neuromodulatory system, biological organisms have the capacity of responding to the ever-changing environment rapidly. This work presents that the mechanism of the neuromodulatory systems through a developmental network can provide a control framework for the artificial agent to regulate its behavior. With the dopamine, serotonin, acetylcholine and norepinephrine modulation, the agent can operate autonomously, effectively carry out specific functions, e.g., to pursue a friend and avoid the enemy, and make suitable and instant decision when the environment changes. Goal-directed pursuing behavior in two simulation scenarios demonstrate the effect of the proposed neural modulatory systems, such as attentional effort, reinforcement learning and addressing the unexpected uncertainty.

We present a framework to analyze chest radiographs for cystic fibrosis using machine learning methods. We compare the representational power of deep learning features with traditional texture features. Specifically, we respectively employ VGG-16 based deep learning features, Tamura and Gabor filter based textural features to represent the cystic fibrosis images. We demonstrate that VGG-16 features perform best, with a maximum agreement of 82%. In addition, due to limited dimensionality, Tamura features for unsegmented images achieve no more than 50% agreement; however, after segmentation, the accuracy of Tamura can reach 78%. In combination with using the deep learning features, we also compare back propagation neural network and sparse coding classifiers to the typical SVM classifier with polynomial kernel function. The result shows that neural network and sparse coding classifiers outperform SVM in most cases. Only with insufficient training samples does SVM demonstrate higher accuracy.

In order to solve the problems such as time consuming and mismatching in the experiment of eliminating SIFT mismatch points in RANSAC algorithm, proposed Mismatching Elimination Algorithm in SIFT Based on Function Fitting; Firstly, we use SIFT algorithm to direct the matching of the image and the matching image, using iterative least squares fitting method to construct function model for the key points of matched Image; secondly, fit the function model with the key points of matching image features; Finally, the errors of the two algorithms are calculated, when the error is greater than the set threshold, verify that the point is a mismatch point, and it is eliminated. The experimental results show that using Mismatching Elimination Algorithm in SIFT Based on Function Fitting than RANSAC algorithm in time to save the 2 s on average, the correct matching rate is increased by 11.75%, and more correct matching points can be reserved.

Sex determination in forensic analysis involves individual examination of different sites of the skull and combination of these sites to understand their impact on the estimation results. Conventionally, forensic experts perform a stepwise combination of several skull region assessment parameters to determine the most important regions with regard to the sex estimation results. This paper introduces a novel group variable selection algorithm: Graph Laplacian Based Group Lasso with split augmented Lagrangian shrinkage algorithm (SALSA) to automatically learn from data by structuring the data into a set of disjointed groups and imposing a number of group sparsity to discover the salient groups which influence the sex determination results. In order to attain this, the skull is partitioned into smaller regions (local regions) using fuzzy c-means (FCM), which are further arranged into clusters as structured groups. Then, we implement the SALSA based group lasso algorithm to impose sparsity on the groups. Our experiments are conducted on 100 skull samples obtained from hospital kuala lumpur (HKL) and the best estimation result obtained is 84.5%.

While data-driven methods for image saliency detection has become more and more mature, video saliency detection, which has additional inter-frame motion and temporal information, still needs further exploration. Different from images, video data, in addition to rich semantic information, also contains a large number of contextual information and motion features. For different scenes, video saliency also has different tendencies. In the movie scene, the face has the strongest visual stimulus to the viewer. In view of the specific movie scene, we propose an efficient and novel video attention prediction model with auxiliary facial saliency (AFSnet) to predict human eye locations in movie scene. The proposed model takes FCN as the basic structure, and improves the prediction effect by adaptively combining facial saliency hints. We give qualitative and quantitative experiments to prove the validity of the model.

Understanding where humans look in a scene is significant for many applications. Researches on neuroscience and cognitive psychology show that human brain always pays attention on special areas when they observe an image. In this paper, we recorded and analyzed human eye-tracking data, we found that these areas mainly were focus on semantic objects. Inspired by neuroscience, deep learning concept is proposed. Fully Convolutional Neural Networks (FCN) as one of methods of deep learning can solve image objects segmentation at semantic level efficiently. So we bring forth a new visual attention model which uses FCN to stimulate the cognitive processing of human free observing a natural scene and fuses attractive low-level features to predict fixation locations. Experimental results demonstrated our model has apparently advantages in biology.

A major challenge for emotion classification using electroencephalography (EEG) is how to effectively extract more discriminative feature and reduce the day-to-day variability in raw EEG data. This study proposed a novel spatial filtering algorithm called Ext-CSP which combined common spatial patterns (CSP) and the regularization term into a unified optimization framework based on Kullback-Leibler (KL) divergence. The experiment was carried out on a five-day Music Emotion EEG dataset of 12 subjects. Four classifiers were applied to make emotion classification. The experiment results demonstrated our unified Ext-CSP algorithm could effectively increase the robustness and generalizability of the extracted EEG features and gain 14% better performance than traditional PCA algorithm, and 1.7% better performance than the stepwise DSA-CSP iteration algorithm on EEG-based emotion classification.

The clonal selection algorithm (CSA), which describes the basic features of an immune response to an antigenic stimulus, has drawn a lot of research attention in the bio-inspired computing community, due to its highly-adaptive and easy-to-implement nature. However, despite many successful applications, this optimization technique still suffers from limited ability to explore the solution space. In this paper, we incorporate the differential evolution (DE) and estimation of distribution algorithm (EDA) into CSA, and thus propose a novel bio-inspired computing algorithm called CSA-DE/EDA. In this algorithm, the hypermutaion and receptor editing processes are implemented based on DE and EDA, which provide improved local and global search ability, respectively. We have applied this algorithm to brain image segmentation. Our comparative experimental results suggest that the proposed CSA-DE/EDA algorithm outperforms several bio-inspired computing techniques on the segmentation problem.

Alzheimer’s disease (AD) has become a nonnegligible global health threat and social problem as the world population ages. The ability to identify AD subjects in an early stage will be increasingly important as disease modifying therapies for AD are developed. In this paper, we propose an ensemble of 3D convolutional neural networks (en3DCNN) for automated identification of AD patients from normal controls using structural magnetic resonance imaging (MRI). We first employ the anatomical automatic labeling (AAL) cortical parcellation map to obtain 116 cortical volumes, then use the samples extracted from each cortical volume to train a 3D convolutional neural network (CNN), and finally assemble the predictions made by well-performed 3D CNNs via majority voting to classify each subject. We evaluated our algorithm against six existing algorithms on 764 MRI scans selected from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. Our results indicate that the proposed en3DCNN algorithm is able to achieve the state-of-the-art performance in early identification of Alzheimer’s Disease using structural MRI.

The traditional classification method based on supervised learning classifies remote sensing (RS) images by using sufficient labelled samples. However, the number of labelled samples is limited due to the expensive and time-consuming collection. To effectively utilize the information of unlabelled samples in the learning process, this paper proposes a novel semi-supervised classification method based on class certainty of samples (CCS). First, the class certainty of unlabelled samples obtained based on multi-class SVM is smoothed for robustness. Then, a new semi-supervised linear discriminant analysis (LDA) is presented based on class certainty, which improves the separability of samples in the projection subspace. Finally, the nearest neighbor classifier is adopted to classify the images. The experimental results demonstrate that the proposed method can effectively exploit the information of unlabelled samples and greatly improve the classification effect compared with other state-of-the-art approaches.

It is of great interest in spectral-spatial features classification for High spectral images (HSI) with high spatial resolution. This paper presents a new Spectral-spatial method for improving accuracy of hyperspectral image classification. Specifically, a new texture feature extraction algorithm based on traditional LBP method is proposed directly. Texture profiles is obtained by the proposed method. A composite kernel framework is employed to join spatial and spectral features. The classifiers adopted in this work is the multinomial logistic regression. In order to illustrate the good performance of the proposed framework, the two real hyperspectral image datasets are employed. Our experimental results with real hyperspectral images indicate that the proposed framework can enhance the classification accuracy than some traditional alternatives.

With the fast development of remote sensing techniques, the spatial resolution of remote sensed image are improved significantly. However, the excessive spatial resolution leads to a sharp increase in data volume and spectral information confusion of objects. The multi-scale kernel learning (MSKL) method has shown an excellent advantage in classification of high-resolution satellite image. Nevertheless, the performance of the MSKL is dramatically influenced by the widths and weights of the Radial Basis Function (RBF) kernel, since its multi-scale kernel function is constructed by several RBF kernels. In order to achieve efficient multi-scale classifier, a new dynamic differential evolution (DE) algorithm is introduced in this paper. In addition, the spectral features and spatial fractal texture features of images are synthetically employed to construct the multi-scale kernel. The experimental results show that the multi-scale kernel based on the dynamic DE algorithm is superior to the traditional multi-scale kernel in obtaining a better multi-scale kernel classifier and with higher classification accuracy.

Recognizing facial expressions is important in many fields such as computer-human interface. Though different approaches have been widely used in facial expression recognition systems, there are still many problems in practice to achieve the best implementation outcomes. Most systems are tested via the lab-based facial expressions, which may be unnatural. Particularly many systems have problems when they are used for recognizing the facial expressions being used during conversation. This paper mainly conducts an experimental study on Eigenface algorithm-based facial expression recognition. It primarily aims to investigate the performance of both lab-based facial expressions and facial expressions used during conversation. The experiment also aims to probe the problems arising from the recognition of facial expression in conversations. The study is carried out using both the author’s facial expressions as the basis for the lab-based expressions and the facial expressions from one elderly person during conversation. The experiment showed a good result in lab-based facial expressions, but there are some issues observed when using the case of facial expressions obtained in conversation. By analyzing the experimental results, future research focus has been highlighted as the investigation of how to recognize special emotions such as a wry smile and how to deal with the interferences in the lower part of face when speaking.

Band selection is of great important for hyperspectral image processing, which can effectively reduce the data redundancy and computation time. In the case of unknown class labels, it is very difficult to select an effective band subset. In this paper, an unsupervised band selection algorithm is proposed which can preserve the original information of the hyperspectral image and select a low-redundancy band subset. First, a search criterion is designed to effectively search the best band subset with maximum information entropy. It is challenging to select a low-redundancy spectral band subset with maximizing the search criteria since it is a NP-hard problem. To overcome this problem, a double-graph model is proposed to capture the correlations between spectral bands with full use of the spatial information. Then, an improved Determinantal Point Process algorithm is presented as the search method to find the low-redundancy band subset from the double-graph model. Experimental results verify that our algorithm achieves better performance than other state-of-the-art methods.

In this work, a dense pyramid network is proposed to provide fine classification maps of high resolution aerial images. The network applied densely connected convolutions to take full advantage of features and deepen the network without concerning the disappearance of gradients. Pyramid pooling module is introduced to bring flexible context information to the segmentation task and accomplish the fusion of multi-resolution features. Additionally, in order to preserve more information of multi-sensor data, group convolutions and channel shuffle operation are applied at the beginning of the network. We evaluate the dense pyramid network on the ISPRS Vaihingen 2D semantic labeling dataset, and the results demonstrate that the proposed framework exhibits better performance compared to the state of the art methods.

Correlation Filter-based trackers have achieved excellent performance and run at high frame rates. Recently, Staple, which utilizing a simple combination of a Correlation Filter (using HOG features) and a global color histogram, has achieved excellent performance. It shows strong robustness in challenging situations including motion blur, illumination changes and deformation changes. However, Staple is only a linear combination of two methods. It is not reliable to determine the confidence level only by the peak. In this paper, we propose Gaussian-Staple that utilize a more sensible way of fusion without destroying the response distribution after fusion. Gaussian prior is added to the response of the output, which is used to determine whether to fine tune by local search. Extensive experiments on a commonly used tracking benchmark show that the proposed method significantly improves Staple, and achieves a better performance than other state-of-the-art trackers.

Many features have been proposed to improve the accuracy of person re-identification. Due to the illumination and viewpoint changes between different cameras, individual feature is less discriminative to separate different persons. In this paper, we propose a saliency-weighted feature descriptor and global-local fusion optimization for person re-identification. Firstly, the weights on pixels are calculated via saliency detection method, then the computed weights are integrated into local maximal occurrence (LOMO) feature descriptor. Secondly, the saliency weights are used to update the metric learning distance in training so that we can learn a new metric matrix for testing. And then, the whole person image is divided into upper and lower halves. A novel global-local fusion method is proposed to combine local and global regions together in the most appropriate way. After that an optimization algorithm is proposed to learn the weights among upper half, lower half and the whole image. According to those weights, a final fused distance is obtained. Experimental results show that the proposed method outperforms many state-of-the-art person re-identification methods.

Kernel extreme learning machine (ELM) has attracted more and more attentions due to its good performance compared with support vector machine (SVM). Since the original Kernel ELM (KELM) is just a spectral classifier, it can’t extract the rich spatial information of hyperspectral images (HSIs). This hence refrains the performance of KELM. In view of this, based on the fact that the neighbors of a pixel are more likely to belong to the same class, this paper proposes a spectral and spatial KELM, which exploits the local spatial information to improve the KELM for HSIs classification. Experimental results on two well-known datasets demonstrate the good performance of the proposed spectral and spatial KELM compared with the original KELM and other state-of-the-art methods.

The huge variance of human pose and inaccurate detection significantly increase the difficulty of person re-identification. Existing deep learning methods mostly focus on extraction of global feature and local feature, or combine them to learn a discriminative pedestrian descriptor. However, rare traditional methods have been exploited the association of the local and global features in convolutional neural networks (CNNs), and some important part-wise information is not captured sufficiently when training. In this paper, we propose a novel architecture called Local-Global Extraction Unit (LGEU), which is able to adaptively re-calibrate part-wise information with integrating the channel-wise information. Extensive experiments on Market-1501, CUHK03, and DukeMTMC-reID datasets achieve competitive results with the state-of-the-art methods. On Market-1501, for instance, LGEU achieves 91.8% rank-1 accuracy and especially 88.0% mAP.

This paper first analysed the state-of-the-art corner detection algorithms and then proposed a novel corner detection approach based on a maximum point-to-chord distance. The proposed corner detector consists of three steps: First, several curves of original image is extracted using Canny edge detector. Second, a method of maximum point-to-chord distance is used in each curve to get the initial corner points. Third, non-maximum suppression and threshold are used to remove corner points with low curvature and get the final result. Different from the CPDA (chord-to-point distance accumulation) corner detector, our proposed detector neither need to accumulate each distance from a moving chord, nor need to computer the accumulation of each point in a curve, therefore achieves better speed while keeping the good average repeatability and accuracy. Compared with the existing methods, the proposed detector attains better performance on average repeatability and localization error under affine transforms, JPEG compression and Gaussian noise.

Due to the distribution of fabric defect shown the sparseness, it is possible to describe the fabric defects feature using sparse representation in particular transform. In this paper, we proposed a novel approach based on sparse representation for detecting patterned fabric defect. In our work, the defective fabric image is expressed by sparse representation model, it is represented as a linear superposition of three components: defect, background and noise. The defective components can be decomposed effectively by using the principle of base pursuit denoising algorithm and block coordination relaxation algorithm. The fabric defect detection is realized by analyzing the defect components. Experimental results demonstrate that the proposed approach is more efficient to detect a variety of fabric defects, in particularly the pattern fabrics.

We propose a novel salient superpixel based tracking algorithm using Coarse-to-Fine segmentation on graph model, where target state is estimated by a combination of pixel-level cues and middle-level cues to achieve accurate target appearance model. We exploit temporal optical flow and color distribution characteristics as coarse grained information from pixel-level processing, and propagate to fine-grained superpixels to improve initial target appearance segmentation from bounding box annotations. Our algorithm constructs a graph model with manifold ranking by improved superpixels to estimate the saliency of target foreground and background in subsequent frames. The tracking result is located by calculating the weight of multi-scale box, where the weight depends on the similarity of scores of foreground and background superpixels in the scale box. We compared our algorithm with the existing techniques in OTB100 dataset, and achieved substantially better performance.

Cross-modal image registration is an intractable problem in computer vision and pattern recognition. Inspired by that human gradually deepen to learn in the cognitive process, we present a novel method to automatically register images with different modes in this paper. Unlike most existing registrations that align images by single type of features or directly using multiple features, we employ the “regenerated” mechanism cooperated with a dynamic routing to adaptively detect features and match for different modal images. The geometry-based maximally stable extremal regions (MSER) are first implemented to fast detect non-overlapping regions as the primitive of feature regeneration, which are used to generate novel control-points using salient image disks (SIDs) operator embedded by a sub-pixel iteration. Then a dynamic routing is proposed to select suitable features and match images. Experimental results on optical and multi-sensor images show that our method has a better accuracy compared to state-of-the-art approaches.

This paper presents a bottom-up saliency model inspired by end-stopping mechanism in primary visual cortex (V1). By modelling an end-stopped cell as multiplication of the outputs from two different orientations tuned selective neurons, corners, line intersections, and line endings, which are called end-stopping features in this paper, are extracted and integrated to indicate saliency cues. The proposed model is constructed as follow: firstly we utilize log-Gabor filters to represent orientation selectivity in V1 neurons; then energy maps of the log-Gabor response from two different orientations are multiplied to extract median features perceived by end-stopped cells; finally the resulting feature maps are combined with color features computed by the traditional center-surround operation to obtain the final saliency map. Results on public eye tracking datasets show the proposed model achieves state-of-the-art performance compared to other models.

To track objects efficiently and effectively in adverse illumination conditions even in dark environment, this paper presents a novel multispectral approach to deploy the intra- and inter-spectral information in the correlation filter tracking framework. Motivated by brain inspired visual cognitive systems, our approach learns the collaborative sparse correlation filters using color and thermal sources from two aspects. First, it pursues a sparse correlation filter for each spectrum. By inheriting from the advantages of the sparse representation, our filers are robust to noises. Second, it exploits the complementary benefits from two modalities to enhance each other. In particular, we take their interdependence into account for deriving the correlation filters jointly, and formulate it as a $${l}_{2,1}$$ -based sparse learning problem. Extensive experiments on large-scale benchmark datasets suggest that our approach performs favorably against the state-of-the-arts in terms of accuracy while achieves in real-time frame rate.

Saliency detection is an important problem in computer vision. Recently, graph-based manifold ranking (GMR) has been successfully employed in image saliency detection problem. Traditional GMR involves two main ranking stages, i.e., ranking with background queries and ranking with foreground queries. However, these two ranking stages are conducted separately which obviously ignores the correlation between background and foreground queries. Also, traditional GMR uses a single graph which lacks of considering multi-view features. To overcome these problems, in this paper, we propose a new multi-view synchronized manifold ranking for saliency detection problem. Our method aims to perform background and foreground ranking simultaneously by exploiting multiple kinds of features and thus performs more robustly and discriminatively for saliency detection problem. Experimental results on three benchmark datasets demonstrate the effectiveness of the proposed saliency detection method.

Sparse representation-based visual tracking methods do not adapt well to changes in the target and backgrounds, and the sparseness of samples does not guarantee optimality. In this paper, we propose a robust visual tracking algorithm using sparse multi-feature selection and adaptive dictionary update based on weight dictionaries. We exploit the color features and texture features of the learning samples to obtain different discriminative dictionaries based on the label consistent K-SVD algorithm, and use the position information of those samples to assign weights to the dictionaries’ base vectors, forming the weight dictionaries. For robust visual tracking, we adopt a novel feature selection strategy that combines the weights of dictionaries’ base vectors and reconstruction errors to select the best sample. In addition, we introduce adaptive noise energy thresholds and establish a dictionary updating mechanism based on noise energy analysis, which effectively reduces the error accumulation caused by dictionary updating and enhances the adaptability to target and background changes. Comparison experiments show that the proposed algorithm performs favorably against several state-of-the-art methods.

Traditional saliency detection via Markov chain only consider boundaries nodes. However, in addition to boundaries cues, background prior and foreground prior cues play a complementary role to enhance saliency detection. In this paper, we propose an absorbing Markov chain based saliency detection method considering both boundary information and foreground prior cues. The proposed approach combines both boundaries and foreground prior cues through bidirectional Markov chain. Specifically, the image is first segmented into superpixels and four boundaries nodes (duplicated as virtual nodes) are selected. Subsequently, the absorption time upon transition node’s random walk to the absorbing state is calculated to obtain foreground possibility. Simultaneously, foreground prior as the virtual absorbing nodes is used to calculate the absorption time and obtain the background possibility. Finally, two obtained results are fused to obtain the combined saliency map using cost function for further optimization at multi-scale. Experimental results demonstrate the outperformance of our proposed model on 4 benchmark datasets as compared to 17 state-of-the-art methods.

Pedestrian detection is a key issue in computer vision, which received extensive attentions. Supervised learning methods with feature extraction and classification are widely used in the pedestrian detection. This paper proposed a pedestrian detection method based on visual saliency and supervised learning. The LC algorithm is used to calculate the saliency value of each training image, followed by the LBP feature extraction. The saliency LBP features and HOG features are combined together as the input of SVM classifier to detect pedestrians. Experimental results show that this method is more accurate and efficient compared with the traditional HOG and LBP feature fusion based method.

Hadoop Distributed File System (HDFS) is designed to reliably storage and manage large-scale files. All the files in HDFS are managed by a single server, the NameNode. The NameNode stores metadata, in its main memory, for each file stored into HDFS. HDFS suffers the penalty of performance with increased number of small files. It imposes a heavy burden to the NameNode to store and manage a mass of small files. The number of files that can be stored into HDFS is constrained by the size of NameNode’s main memory. In order to improve the efficiency of storing and accessing the small files on HDFS, we propose Small Hadoop Distributed File System (SHDFS), which bases on original HDFS. Compared to original HDFS, we add two novel modules in the proposed SHDFS: merging module and caching module. In merging module, the correlated files model is proposed, which is used to find out the correlated files by user-based collaborative filtering and then merge correlated files into a single large file to reduce the total number of files. In caching module, we use Log - linear model to dig out some hot-spot data that user frequently access to, and then design a special memory subsystem to cache these hot-spot data. Caching mechanism speeds up access to hot-spot data.The experimental results indicate that SHDFS is able to reduce the metadata footprint on NameNode’s main memory and also improve the efficiency of storing and accessing large number of small files.

Steganography is a kind of technique which hides data under the cover file so as not to arouse any suspicion. In this paper, a new image steganography algorithm combining compressive sensing (CS) and random embedding is proposed. There are two security parts in this algorithm. The first part is the random projection inherited from CS, and the second is the random embedding process. CS serves to create the encrypted data, and acts as a tool to reduce the dimensionality of the secret data. Random-embedding algorithm is proposed to choose the position of cover image randomly for hiding the secret image. This algorithm uses symmetric key method, which means the sender and the receiver use the same key. Numerical experiments show that this steganography algorithm provides high embedding capacity and high Peak Signal Noise Ratio (PSNR).

With the development of abnormal behavior analysis technology, measuring the similarity of abnormal behavior has become a core part of abnormal behavior detection. However, there are general problems of central selection distortion and slow iterative convergence with existing clustering-based analysis algorithms. Therefore, this paper proposes an improved clustering-based abnormal behavior analysis algorithm by using K-means. Firstly, an abnormal behavior set is constructed for each user from his or her behavioral data. A weight calculation method for abnormal behaviors and an eigenvalue extraction method for abnormal behavior sets are proposed by using all the behavior sets. Secondly, an improved algorithm is developed, in which we calculate the tightness of all data points and select the initial cluster centers from the data points with high density and low density to improve the clustering effect based on the K-means clustering algorithm. Finally, clustering result of the abnormal behavior is got with the input of the eigenvalues of the abnormal behavior set. The results show that, the proposed algorithm is superior to the traditional clustering algorithm in clustering performance, and can effectively enhance the clustering effect of abnormal behavior.

Adaptive Lasso preserves oracle properties comparing to classical Lasso. It performs as well as if the true underlying model is provided in advance. In order to let feature subset selected by Adaptive Lasso preserve more local information, which is discriminative and benefit for classification, Manifold-regularized Adaptive Lasso (MrALasso) is proposed for feature selection. Reconstructing response by linear sum of features is considered in manifold embedded in high-dimensional space. A similarity graph of data points is built. Connected points are restricted to stay together as close as possible so that the intrinsic geometry of the data and the local structure are preserved. An effective iterative algorithm, with detailed proof of convergence, is proposed to solve the optimization problem. Experimental results of feature selection on several classical gene datasets show the effectiveness and superiority of the proposed method.

Data annotation is an important but time-consuming and costly procedure. To sort a text into two classes, the very first thing we need is a good annotation guideline, establishing what is required to qualify for each class. In the literature, the difficulties associated with an appropriate data annotation has been underestimated. In this paper, we present a novel approach to automatically construct an annotated sentiment corpus for Algerian dialect (A Maghrebi Arabic dialect). The construction of this corpus is based on an Algerian sentiment lexicon that is also constructed automatically. The presented work deals with the two widely used scripts on Arabic social media: Arabic and Arabizi. The proposed approach automatically constructs a sentiment corpus containing 8000 messages (where 4000 are dedicated to Arabic and 4000 to Arabizi). The achieved F1-score is up to 72% and 78% for an Arabic and Arabizi test sets, respectively. Ongoing work is aimed at integrating transliteration process for Arabizi messages to further improve the obtained results.

Automatic story segmentation is an important prerequisite for semantic-level applications. The normalized cuts (NCuts) method has recently shown great promise for segmenting English spoken lectures. However, the availability assumption of the exact story number per file significantly limits its capability to handle a large number of transcripts. Besides, how to apply such method to Chinese language in the presence of speech recognition errors is unclear yet. Addressesing these two problems, we propose a self-validated NCuts (SNCuts) algorithm for segmenting Chinese broadcast news via inaccurate lexical cues, generated by the Chinese large vocabulary continuous speech recognizer (LVCSR). Due to the specialty of Chinese language, we present a subword-level graph embedding for the erroneous LVCSR transcripts. We regularize the NCuts criterion by a general exponential prior of story numbers, respecting the principle of Occam’s razor. Given the maximum story number as a general parameter, we can automatically obtain reasonable segmentations for a large number of news transcripts, with the story numbers automatically determined for each file, and with comparable complexity to alternative non-self-validated methods. Extensive experiments on benchmark corpus show that: (i) the proposed SNCuts algorithm can efficiently produce comparable or even better segmentation quality, as compared to other state-of-the-art methods with true story number as an input parameter; and (ii) the subword-level embedding always helps to recovering lexical cohesion in Chinese erroneous transcripts, thus improving both segmentation accuracy and robustness to LVCSR errors.

Deep Learning has been considered as an effective tool for Big Data Analytics due to its capabilities of dealing with massive amounts of complex structured and unstructured data. Deep Learning has recently come to play a significant role in solutions for Big Data Analytics. The Sentiment Analysis is also considered the most effective tool for performing the real-time analytics to know “what is really happening now” queries.This paper studies the method that integrated the Deep Learning Model with a Real-Time Sentiment Analysis technique to perform predictive analytics that could improve the outcomes of the Big Data Analytics solution for an informed decision-making process. A proof of concept project on Stock Market Prediction System was developed to demonstrate the real value of our approach for an improved Big Data Analytics solution.

In the literature, limited work has been conducted to develop sentiment resources for Saudi dialect. The lack of resources such as dialectical lexicons and corpora are some of the major bottlenecks to the successful development of Arabic sentiment analysis models. In this paper, a semi-supervised approach is presented to construct an annotated sentiment corpus for Saudi dialect using Twitter. The presented approach is primarily based on a list of lexicons built by using word embedding techniques such as word2vec. A huge corpus extracted from twitter is annotated and manually reviewed to exclude incorrect annotated tweets which is publicly available. For corpus validation, state-of-the-art classification algorithms (such as Logistic Regression, Support Vector Machine, and Naive Bayes) are applied and evaluated. Simulation results demonstrate that the Naive Bayes algorithm outperformed all other approaches and achieved accuracy up to 91%.

The rise of social media is enabling people to freely express their opinions about products and services. The aim of sentiment analysis is to automatically determine subject’s sentiment (e.g., positive, negative, or neutral) towards a particular aspect such as topic, product, movie, news etc. Deep learning has recently emerged as a powerful machine learning technique to tackle a growing demand of accurate sentiment analysis. However, limited work has been conducted to apply deep learning algorithms to languages other than English, such as Persian. In this work, two deep learning models (deep autoencoders and deep convolutional neural networks (CNNs)) are developed and applied to a novel Persian movie reviews dataset. The proposed deep learning models are analyzed and compared with the state-of-the-art shallow multilayer perceptron (MLP) based machine learning model. Simulation results demonstrate the enhanced performance of deep learning over state-of-the-art MLP.

Crime analysis and prediction is a systematic approach for analyzing and identifying different patterns, relations and trends in crime. In this paper we conduct exploratory data analysis to analyze criminal data in San Francisco, Chicago and Philadelphia. We first explored time series of the data, and forecast crime trends in the following years. Then predicted crime category given time and location, to overcome the problem of imbalance, we merged multiple classes into larger classes and did feature selection to improve accuracy. We have applied several state-of-the-art data mining techniques that are specifically used for crime prediction. The experimental results show that the Tree classification models performed better on our classification task over k-NN and Naive Bayesian approaches. Holt-Winters with multiplicative seasonality gives best results when predicting crime trends. The promising outcomes will be beneficial for police department and law enforcement to speed up the process of solving crimes and provide insights that enable them track criminal activities, predict the likelihood of incidents, effectively deploy resources and make faster decisions.

Sentiment analysis mainly focused on the automatic recognition of opinions’ polarity, as positive or negative. Nowadays, sentiment analysis is replacing the web-based and traditional survey methods commonly conducted by companies for finding the public opinion about their products and services to improve their marketing strategy and product advertisement and help to improve customer service. The online availability of large text makes it important to be analyzed. The automatic analysis of this information involves a deep understanding of natural languages. Sentiments and emotions play a pivotal role in our daily lives. They assist decision-making, learning, communication, and situation awareness in human environments. The importance of processing and understanding dialect text is increasing due to the growth of socially generated dialectal content in social media. In addition to existing materials such as local proverbs, advice and folklore that are found spread on the web. This paper focused on text sentiment analysis as dialect text, as quick review to identify relevant contributions that address languages aspect for a specific dialect.

The application of intelligent question answering system in college assistant teaching is an effective way to reduce the workload of university teachers and improve students’ learning efficiency. With the rapid development of related technologies, the intelligent question answering system has made great progress, but there is little related work in answering students’ university course questions, and there are some problems such as poor accuracy and non universality. Because of this reason, it cannot fully meet the demands of universities. Therefore, this paper proposes an intelligent question answering system for professional questions. First, we select candidate question-and-answer pairs in the knowledge base through professional word matching, and then use the attention mechanism proposed in this paper and bi-directional long short term memory network (BiLSTM) to calculate the semantic similarity between query questions and candidate questions. Multiplying the semantic similarity by the keywords similarity of the two questions as the final similarity. Finally, we push the three most similar candidate questions and the corresponding answers to students. The experimental results show that the system improves the accuracy of answering students’ university course questions, and is applicable to any university course.

To make full use of patent information and help companies find similar patent pairs by calculating the similarity of patents, help them deal with the issue of patent infringement detection, patent search, enterprise competition analysis, and patent layout, this paper proposes a method for calculation of patent similarity based on patent text using patent model tree. This method not only simplifies the process of understanding the patent text but also increases the accuracy of calculating the similarity among patents effectively. In this paper, the similarity between patents is calculated based on the patent model tree, and different similarity calculation methods are used according to different properties of tree nodes. Among them, in order to improve the accuracy of the claims node similarity measurement results, the Siamese LSTM network is applied. The experimental results show that the patent similarity calculation method based on text has an outstanding accuracy.

The rapid development of the Internet, especially mobile Internet, makes it much easier for people to make social contacts online. Nowadays people tend to spend more and more time on social network service, and produce a lot of image files. This brings a challenge to traditional standalone framework on handing the continued increasing image files. Therefore, it is advisable to find a new way to settle the challenge. Hadoop is a notable, widely-used project for distributed storage and computations with high efficiency, data integrity, reliability and fault tolerance. Hadoop Distributed File System and MapReduce are two primary subprojects respectively for big data storage and computations. However, Hadoop does not provide any interface for image processing. Moreover, both Hadoop Distributed File System and MapReduce have trouble in processing large amount of small files, which result in decreasing efficiency of files access and distributed computations. This prevents us from performing images processing actions on Hadoop. In view of this, this paper proposes a new method to optimize the storage of small image files on Hadoop and self-defines an input/output format to enable Hadoop to process image files.

Big data analysis and application is an efficient approach for analyzing and identifying different patterns, relations and trends in daily life. In this paper we proposed an intelligent big data platform for information sharing in connection to administrative law and criminal justice. We first explored the structure of the data and utilized Apache Pig and Hadoop to handle structured, semi-structured and unstructured data. We extracted and transformed useful features from data and delivered and stored them in database using Cassandra and Zookeeper. After obtaining required features we applied machine learning and neural network algorithms in the data sets, to classify or mine potential knowledge. Finally we stored the results in MongoDB in which all staff from law enforce departments can have access to them through Web APP. We have applied several state-of-the-art data mining techniques and big data analytic tools that are specifically used for data processing and feature extraction. The experimental results show that our system is efficient in improving the filing rate, and also time-saving in processing large number of data. The promising outcomes will be beneficial for administrative enforcement and law enforcement to speed up the process of solving law cases and provide insights that enable them track case activities, predict the likelihood of warnings, effectively deploy resources and make faster decisions.

In recent research, geometric attack is one of the most challenging problems in digital watermark. Such attacks are very simple to defeat most of the existing digital watermark algorithms without destroying watermark itself. In this paper, a point matching measure is adopted for estimating the geometric transformation parameters. First, the affine invariant points of the original and probe image are computed. Then, the best embedded coefficients are found via GA in which the fitness function is defined as the minimal change of the significant region after the watermarking embedding. Finally, the watermark embedding and extraction were implemented in digital wavelet transform (DWT) domain. The propose scheme actualizes blind extraction since not requiring the original image information. The watermark is embedded adaptively according to image texture. This method has been proved its robustness to various attacks through experiments, and it can recover the watermarking image when the watermarking is aggressed.

VPN is widely used in colleges and universities at present, having brought great convenience to teachers and students in their study and life. Due to the fact that the current VPN in colleges and universities has the same access rights to all users, as long as VPN users login successfully via VPN account, they can access the campus network. As a result, not only is the management of the VPN administrators troublesome, but also the security of the network environment in the school is not very good. In view of this problem, this paper proposes a method that people visit the campus network resources according to user’s access authority in which the distribution of access rights can be distributed according to user’s identity category or individual way, and gives the application in college campus network. Finally, the simulation results show that this system is feasible to guarantee the security of communication.

The redundancy of JPEG images is lower than that of non-compressed images, so any modification will greatly reduce the visual effect and produce file expansion. In this paper, a new RDH algorithm for JPEG images based on interval correlation is proposed. Through the quantization of DCT, the secret message is embedded into a continuous interval with large correlation and low distortion by using histogram shifting (HS). Experimental result shows that the proposed algorithm has better peak signal noise ratio (PSNR) and file expansion than the state-of-the-art HS algorithms for JPEG.

Privacy preservation in a data-sharing computing environment is becoming a challenging problem. A purpose is defined as the intention of data accesses or usages and purpose-based access control (PBAC) has been proposed to extend traditional models for privacy-preserving. However, in existing research, the purpose is often to bind to data by using labeling schemes or building privacy metadata databases separately, which leads to redundancy of data tables and decreasing of query efficiency. Moreover, privacy policies involving purpose lack sufficient semantics. In this paper, we present a semantic model for the role-involved conditional purpose-based access control (RCPBAC) with ontology. Purpose, data, and role are represented by ontology and their relationships are described with object or data properties, which is based on Web Ontology Language (OWL). We use Semantic Web Rule Language (SWRL) to represent privacy policies for reasoning. This model can help data providers to define and share their own information more easily and securely.

In this paper, we provide a real-time image deformation method based on Locally-weighted Moving Least Squares (LW-MLS). To achieve a detail-preserving and realistic deformation of images, a concise deformation formula is proposed as the deformation function. Compared with two state-of-the-art methods, Moving Least Squares (MLS) and Moving Regularized Least Squares (MRLS), the main improvement of our method is preprocessing the control points, which adopts sparse approximation to achieve a fast deformation. With the traditional methods of image deformation, each pixel is affected by all control points, which consume too much time to deform an image. So in our method, each pixel is mainly affected by surrounding control points, and every pixel is almost not affected by the control points which are far away from the deformed pixel. The novel method we proposed can be performed in real time and could supply promising performance for the deformation of large image.

With the rapid development of cloud computing, cloud security is increasingly an important issue. Virtual machine (VM) is the main form to provide cloud service. To protect VMs against malware attack, a cloud needs to have the ability to react not only to known malware, but also to the new emerged ones. Virtual Machine Introspection (VMI) is a good solution for VM monitoring, which can obtain the raw memory state of the VM at Virtual Machine Monitor (VMM) level. Through analyzing the memory dumps, the significant features of malware can be obtained. In our research, we propose a novel static analysis method for unknown malware detection based on the feature of opcode n-gram of the executable files. Different feature sizes ranging from 2-gram to 4-gram are implemented with the feature length of 100, 200, 300 respectively. The feature selection criterion of Term Frequency (TF)-Inverse Document Frequency (IDF) and Information Gain (IG) are leveraged to extract the top features for classifier training. Different classifiers are trained with the preprocessed dataset. The experimental results show that the weighted integrated classifier with opcode 4-gram of 300 features has the optimal accuracy of 98.2%.

Trusted Connection Authentication (TCA) is a critical part of network security access solution. TCA is a kind of high level trusted network access techniques which can create trusted connections between client and remote networks through two-way user authentication and platform identification in TTP. However, there are general security problems after accessed to the network which are not much considered by existing TCA schemes. Therefore, this paper proposes a reinforced TCA architecture, TCA-BA, which extends a network behavior layer on the basis of TCA. Firstly, network behavior eigenvalue extraction is proposed by using time and host network flow characteristics. Secondly, a new method is illustrated in which we classify the behavior by Naïve Bayes Algorithm, measure the network abnormal behavior by minimum risk bayes rules, identify these behaviors which have accessed to the network. Finally, the experimental results present that our architecture can effectively identify the abnormal behavior in the network and protect the network security.

In 5G mobile network, SDN/NFV as a key technology is widely used in EPC networks. In order to cope with the increasing data service in the EPC of 5G network, we propose a proactive cache strategy based on the deep learning network SSAEs for content popularity prediction based on the SDN/NFV architecture, SNDLPC. Firstly, NFV/SDN technique is used to build a virtual distributed deep learning network SSAEs. Then, the SSAEs network parameters are unsupervised trained by the historical users’ data. Finally, the content popularity is predicted by SSAEs using the data of user request in whole network collected by SDN controller. The SDN controller generates the proactive caching strategy according to the prediction results and synchronizes it to each cache node through flowtable to implement the strategy. In the simulation, the SSAEs network structure parameters are compared and determined. Compared with other strategies, such as the typical Hash + LRU and Betw + LRU caching strategies, SVM prediction and the BPNN prediction algorithm, the proposed SNDLPC proactive cache strategy can significantly improve cache performance.

The hand-eye calibration problem is to compute the relative pose between a robot platform (hand) and a camera (eye) mounted rigidly on the robot platform. To solve the problem, the motion pairs of the robot platform movement and the corresponding camera movement are collected and then use a linear algorithm or nonlinear minimization algorithm to find the optimal solution from the collected motion pairs. Because there are noises in the motion pairs, the previous method uses the motion pairs directly can’t effectively reduce the impact of the noise. In this paper, we focus on how to ease the impact of the noises of the motion pairs. We use active vision approach to hybrid the robot platform movement and camera movement with different credibility and generate a series of special motion pairs, this can dynamically get a more accurate initial estimation of the relative pose between the robot platform and the camera. Then we use a recursive way to filter matching points and update the estimation of the relative pose, which can improve the robustness of our hand-eye calibration algorithm effectively. Both virtual and real experiments show the superiority of our approach over the previous methods.

Attackers have developed ever more sophisticated and intelligent ways to hack information and communication technology (ICT) systems. The extent of damage an individual hacker can carry out upon infiltrating a system is well understood. A potentially catastrophic scenario can be envisaged where a nation-state intercepting encrypted financial data gets hacked. Thus, intelligent cybersecurity systems have become inevitably important for improved protection against malicious threats. However, as malware attacks continue to dramatically increase in volume and complexity, it has become ever more challenging for traditional analytic tools to detect and mitigate threat. Furthermore, a huge amount of data produced by large networks have made the recognition task even more complicated and challenging. In this work, we propose an innovative statistical analysis driven optimized deep learning system for intrusion detection. The proposed intrusion detection system (IDS) extracts optimized and more correlated features using big data visualization and statistical analysis methods, followed by a deep autoencoder (AE) for potential threat detection. Specifically, a preprocessing module eliminates the outliers and converts categorical variables into one-hot-encoded vectors. The feature extraction module discards features with null values grater than 80% and selects the most significant features as input to the deep autoencoder model trained in a greedy-wise manner. The NSL-KDD dataset (an improved version of the original KDD dataset) from the Canadian Institute for Cybersecurity is used as a benchmark to evaluate the feasibility and effectiveness of the proposed architecture. Simulation results demonstrate the potential of our proposed IDS system for improving intrusion detection as compared to existing state-of-the-art methods.

Games are both a way to enjoy leisure time and to learn. Understanding how mental processes associated with gaming work at a deeper level is very important, especially with emerging technologies such as consumer VR head-mounted display systems. One approach to better understand games is through the analysis of how individual events and components of the game affect our autonomic responses. To this end, in this paper, we analyze how the component of viewing perspectives and display types affect the reaction to specific in-game events. We do this through the collection of EEG data using a consumer EEG headset. The collected values are used to calculate Arousal-Valence and Engagement indexes. Finally, these values are compared to events happening at the collection time, and the data is analyzed to identify patterns and draw conclusions from the data. This initial analysis of selected events does not identify any representative change in values amongst different displays and viewing perspectives. These results suggest that viewing perspective and display are of less importance than may be expected for our selected events, whereas other factors such as ranking play a greater role in emotional state changes.

This paper introduces a novel Loop subdivision method, which produces a C1 continuity surface including boundaries and creases. The new rules develop Loop subdivision surface by adding a parameter known as a knot interval. Sederberg et al. used knot intervals for sharp features in subdivision surface modeling for the first time. This paper extends the subdivision rule to triangular subdivision meshes. It can generate a pleasant result in Loop subdivision surfaces.

In order to extend the abilities of current robots in industrial applications towards more autonomous and flexible manufacturing, this work presents an integrated system comprising real-time sensing, path-planning and control of industrial robots to provide them with adaptive reasoning, autonomous thinking and environment interaction under dynamic and challenging conditions. The developed system consists of an intelligent motion planner for a 6 degrees-of-freedom robotic manipulator, which performs pick-and-place tasks according to an optimized path computed in real-time while avoiding a moving obstacle in the workspace. This moving obstacle is tracked by a sensing strategy based on machine vision, working on the HSV space for color detection in order to deal with changing conditions including non-uniform background, lighting reflections and shadows projection. The proposed machine vision is implemented by an off-board scheme with two low-cost cameras, where the second camera is aimed at solving the problem of vision obstruction when the robot invades the field of view of the main sensor. Real-time performance of the overall system has been experimentally tested, using a KUKA KR90 R3100 robot.

Color-guided depth image restoration is an issue of great interest. However, the edge in color image is not always consistent with the depth image. There is a certain relationship between the shading component of RGB image and the depth, so a depth image restoration method is proposed with shading structure guidance. First, the RGB image is decomposed into the shading component and the reflectance component based on Retinex Theory; next, calculate the structure tensors of the shading component and the depth image respectively, and the corresponding eigenvalues and eigenvectors; then, design the diffusion tensor with the eigenvalues and eigenvectors of the depth structure tensor to make the diffusion be along the level lines isophotes, finally the shading structure is introduced to inhibit the diffusion in the direction perpendicular to the edge, and the depth image is restored by diffusion. Experiments show, visually and quantitatively, the better restoration results are achieved by the introduction of the shading structure.

Muon imaging is a new imaging technique which can be used to image large bulky objects, especially objects with heavy shielding where other techniques like X-ray CT scanning will often fail. This is due to the fact that high energy cosmic rays have a very high penetrative power and can easily penetrate hundreds of meters of rock. Muon imaging is essentially an inverse problem. There are two popular forms of muon imaging techniques - absorption muon imaging based on the attenuation of muons in matter and multiple scattering muon imaging based on the multiple scattering effect of muons. Muon imaging can be used in many areas, ranging from volcanology and searching for secret cavities in pyramids over border monitoring for special nuclear materials to nuclear safeguards applications for monitoring the spent fuel casks. Due to the lack of man-made muon sources, both of the muon imaging techniques rely on cosmic ray muons. One important shortcoming of comic ray muons are the very low intensities. In order to get high image resolutions, very long exposure times are needed. In this paper, we will study how machine learning techniques can be used to improve the muon imaging techniques.

In recent years, image processing has been applied in various industries. In the part of the public road scene, the vehicle camera in night could not be used perfectly as it does in daytime, because it usually gains low visibility by the faint illumination. In order to enhance the visual clear visibility, in this paper, the mixed multi-Retinex algorithm is first introduced to deal with the night view scene, and then using fractional differentiation to make the edge information much clearer. Finally we combined the two methods with center surround function to make the effects better. In the experiment, processing speed is faster than the comparing methods and dark regions has boosted the brightness of the image.

The image of Highway Traffic defogging has become an important part in the traffic security monitoring system. The studies in the image field on this topic are mainly based on the optical image transform theory and mathematical physics, with a number of fruitful results have been achieved, this paper proposes a method combining the mathematical morphology and the bit plane decomposition coding for traffic image defogging, compared with the traditional image defogging methods, the new method makes the effect increased displayed on PSNR values.

The simulation of a non-Gaussian scattering on rough surface based on local curvature approximation (NG-LCA) model is presented. The comparison between the NRCS result of LCA and the QuikSCAT scatterometer data shows that NG-LCA model can well explain the scattering way of the Upwind/downwind asymmetry.

This paper presents a fuzzy logic control for a distributed multi-node temperature control. The fuzzy logic controller is also introduced to the system for keeping temperature index to be constant. Because real-time induction temperature has many differences with correlation in industry. So the temperature control of induction is more important to control temperature of each node. The result of main fuzzy controller and five node fuzzy controllers of temperature will be conducted in this paper. It is observed that the effect of adjacent node fuzzy controller on performance of distributed multi-node temperature. The node of adjacent fuzzy controller can avoid strong sway phenomenon, and industrial temperature control system introduced in paper can get good regulation quality.

Aiming at the problem of the slow speed of reinforcement learning, a tentative Q learning algorithm is proposed. By improving the number of exploration in each learning iteration and the updating method of Q table, tentative Q learning algorithm accelerates the learning speed and ensures the balance between exploration and exploitation. Finally, the feasibility and effectiveness of the algorithm are proved by the experiment of robot path planning.