Abstract:Deep learning has recently achieved breakthrough progress in speech recognition and image recognition. With the advent of big data era, deep convolutional neural networks with more hidden layers and more complexarchitectures have more powerful ability of feature learning and feature representation. Convolutional neural network models trained by deep learning algorithm have attained remarkable performance in many large scale recognition tasks of computer vision since they are presented. In this paper, the arising and development of deep learning and convolutional neural network are briefly introduced, with emphasis on the basic structure of convolutional neural network as well as feature extraction using convolution and pooling operations. The current research status and trend of convolutional neural networks based on deep learning and their applications in computer vision are reviewed, such as image classification, object detection, pose estimation, image segmentation and face detection etc. Some related works are introduced from the following three aspects, i.e., construction of typical network structures, training methods and performance. Finally, some existing problems in the present research are briefly summarized and discussed and some possible new directions for future development are prospected.

Abstract:With the rapid development of information technology and social media, The travel information increases. The people′s travel demands and preferences are gradually diversified and customized, so tourism information service has become a hot topic among researchers. Moreover, communities and user-operated media in social media have tremendous tourism information resources which can be employed in travel information mining based on social media. Since travel information is used for smart travel applications, it promotes the fast growing of internet-based tourism informatization. In this paper, the background of internet-based tourism inforamatization is discussed. Secondly, the characteristics of travel information in social media are analyzed. Then, some hot research topics in tourism applications are presented and discussed. Finally, the challenges of social media based travel data mining and applications are summarized, and some further research directions are explored.

Abstract:Magnetic resonance imaging (MRI) has several advantages over other medical imaging modalities, including high contrast among different soft tissues, relatively high spatial resolution across the entire field of view and multi-spectral characteristics. Hence, it has been widely used in quantitative brain imaging studies. Quantitative volumetric measurement and three-dimensional visualization of brain tissues are helpful for pathological evolution analyses, where image segmentation plays an important role. However, MR images suffer from several major artifacts, including intensity inhomogeneity, noise, partial volume effect and low contrast, which makes MR segmentation remain a challenging topic. Therefore, this paper reviews brain MR image segmentation based on fuzzy clustering model from seven aspects, i.e., the determination of cluster number, the initialization of model, the robustness to noise, the estimation of intensity inhomogeneity and partial volume, the uncertainty description of data and the model extension. Limitations existing in the available methods are analyzed, and problems in further research are discussed as well.

Abstract:In recent years, research on automatic pain recognition is of increased attetion , due to its wide application in clinics, especially for the treatment and nursing of patients who cannot express their pain vocally. Since the face is the vital cue for evaluating pain and the great progress has been made in facial expression recognition with computer vison technique, it is an effective way to recognize pain automatically utilizing facial information. Here, four existing databases used for pain recognition are firstly introduced, namely, the STOIC database, the Infant COPE database, the UNBC-McMaster Shoulder Pain Expression Archive database, and the BioVid Heat Pain databse. Then, the proposed methods in the last decade can be divided into four categories depending on the use of either static images, video sequences, person specific strategy or multimodal methods. Finally, the current state of the art in pain detection research, open issues and future directions are highlighted.

Abstract:Software quality ensures the reliable running of the software system, and soft ware defects reduce the quality of the software system. Software defects can be identified effectively by mining the codes as well as other related data, so the software defect mining technology has drawn significant attention in software quality assurance. To effectively identify potential software defects from the software modules, a large number of modules labeled as defective or non-defective information need to be collected for model construction. However, the labels of modules are usually obtained by extensive testin g or manual code inspection, which consumes a huge amount of manpower and time. In pract ice, only a small number of labels can be collected, which seriously constrains the performance of defectidentification. To solve this problem, the semi-supervised learning is introduced into software defect mining, thus the mining performance is improved by exploiting the large number of unlabeled modules. Here, the advances and the research status of semi-supervised software defect mining are reviewed and discussed extensively. Firstly, the existing studies on software defect mining is briefly review, and then the four major paradigms of semi-supervised learning are introduced. Finally, various methods and techniques on semi-supervised defect mining are systematically summarized and reviewed.

Abstract:License plate localization, the core component of license plate recognition system, is valuable in both academic development and potential applications. Though much progress has been made in recent years, challenging problems still exist in the complex scenes, such as low luminance, low resolution and inclination scence of vehicle. This paper proposes a novel fully convolutional neural network to localize license plates accurately by a corner regression algorithm. To guarantee effective training in the proposed model, 45 000 sample images are annotated by one person. Meanwhile, the annotated sample images are processed by four operators, including translating, scaling, rotating and noising, to increase the number and diversity of the training samples. Extensive experiments on the newly collected datasets, trafficmonitoring dataset and the complex scene dataset, demonstrate the effectiveness of the proposed method against other two license plate localization methods.

Abstract:To solve the contradictory of speed and accuracy of the existing interpolation method, an interpolation method based on sub-regional and multi-directional data fusion is presented. The new method divides images into flat areas and edge-regions. The bilinear algorithm is used in flat areas, and the improved algorithm is used in edge-regions. Based on the inverse ratio of the distance square, the improved algorithm fully considers four-direction estimated results of the nearest horizontal, vertical and two diagonal directions using 12 pixels selected from 4×4 neighborhood. A final interpolation result is obtained by calculating weights of vertical distance and direction gradient. The experimental result shows that the proposed interpolation method costs less time and can make edges of image more natural and clear. Moreover, it can be applied to any multiples of interpolation amplification.

Abstract:For wireless multimedia sensor networks, in the radio channel there exist random fading, high error rate and other issues, especially on video applications. Applying the compressed sensing theory to video encoding provides an approach to resist random wireless channel fading and reduce error rate. However, the high complexity of compressed sensing reconstruction algorithm makes it difficult to recover the real-time video sequences at the decoder. Here, by improving iterative search algorithm, iterative search method and the loop termination conditions, a fast and efficient smoothed l0 norm-based video decoding compressed sensing algorithm (ADSL0) is presented. To ensure the optimal iteration path, the algorithm uses a strict descent direction and correction iterative step. Experimental results show that the proposed algorithm in terms of time-consumingre modeling or reconstruction accuracy is significantly better than other similar algorithms.

Abstract:A method of to hydrological sheet color image segmentation based on gradient and color information is proposed and applied to paper hydrology data digitization to deal with the hydrological sheet images taken by camera. Firstly, curves are obtained by making use of color feature in the CIE Lab color space. Then the image is processed in a block-by-block manner. The gradient operators in horizontal and vertical directions are performed to discriminate the target pixels on the grid lines roughly. Thirdly, the color information of those pixels is obtained and the threshold to segment the grid lines are set. After getting the grid lines, horizontal and verticalerosions are implemented and most of the noise is removed. Finally, the segmented curves and the grid lines are merged and the final segmentation is established. The experimental results on several hydrological sheet color images show that the proposed method can fulfill the goal of image segmentation effectively in a self-adaption way and alleviate the effect of uneven illumination with good robustness and lower computational complexity.

Abstract:Redundancy of the multi-subspaces′ fusion data represent by features can be minimized with the direct summation over multi-subspaces. In this paper, a new face recognition method based on feature fusion was proposed via using the direct summation of multi-subspaces. First we calculate the covariance matrices of all training samples′front face, left face and right face images, which are all normalized, and then calculate their first Plargest eigenvalues and corresponding mutually orthogonal eigenvectors, using the 2DPCA algorithm. Then we constitute three feature space (projection space) via three multi-subspaces′ orthogonal eigenvectors which meet the direct sum condition. The samples′ front face, left face and right face images are projected into the three spaces respectively. The projected features are fused as the classification feature. The comparison on the three groups of experimental results shows that our algorithm not only reduce the computation but also increase the recognition rate.

Abstract:Video synopsis is a temporally compact representation of the original video, which facilitates the subsequent video processing, such as video storage, browsing and retrieval. Most of conventional methods easily lose some important objects and can not represent the original videos completely. Therefore, this paper proposes a novel method based on object trajectory optimization. The method extracts object trajectories using an improved multi-object tracking method, and optimizes the temporal shift labels of those trajectories. The optimal labels are then formulated as the maximum a posteriori state of a special Markov random field, which can be solved by the relaxed linear programming method. The synopsis video is obtained by integrating the optimal labels into the background sequence. Extensive experiments on both public and collected video sequences suggest that our method outperforms other methods in accuracy. In particular, our method can retain most essential information of the video sources in shorter synopsis videos.

Abstract:To investigate the co-movement of stock, traditional time series analysis and data mining technology mainly use domestic or foreign stock index to study the co-movement bet ween market, sector or industry, and obtain some macroscopic conclusion. Therefore, there is a lack of direct analysis and mining linkage between individual stocks data issues. A method based on dynamic time warping is proposed to analyze the co-movement between two individual stocks. It can find some similar stocks in shape and obtain relevant essential information from extra-large stocks. Combining with k-means clustering method based on dynamic time warping, the clustering method can gain some clusters which have the same fluctuation tendency. The results demonstrate that the proposed method can accurately find the stocks which have linkage relationship from large amounts of stocks, as well as separating clusters of different fluctuation of stocks. It shows that the proposed method has a certain superiority.

Abstract:Sequential information bottleneck (sIB) algorithm is one of the widely used clustering algorithms. The sIB algorithm applies the joint probability model to describe data, which has good ability to express the relationship between data samples and data attributes. However, the sIB algorithm suggests that all data attributes are equally important, which influences the clustering effect. To address the issue, the paper proposes the weighting joint probability model. The proposed model applies the mutual information measurement to the important level of data attributes so that to highlight representative attributes and depress redundancy attributes. Experiments on UCI datasets show that the proposed the weighting joint probability model (WJPM) sIB algorithm based on WJPM improves the F1 measure by 5.90% than the sIB algorithm.

Abstract:Principal component analysis (PCA) is the well-known method in pattern recognition. However, expanding original image matrices into the same dimensional vectors in classical PCA increase the computational complexity. Here one presents a kind of multi-band principle component analysis (MBPCA). The process can reduce the computational complexity thus improving the overall performance. Firstly, the image is transformed into frequency data by the two-dimensional discrete cosine transform. Secondly, frequency data is divided into a plurality of frequency bands according to its frequency range. Finally, a principal component analysis method using a plurality of frequency bands is designed. The experiments on ORL and NUST603 face database show that the proposed method has the ability to quickly extract image features and performs better than the corresponding principal component analysis.

Abstract:To locate payloads for the least significant bit matching (LSB-M) steganography, the paper proposes a new method. The problem of payload location for LSB-M can be solved by abstracting the mean square adjacency pixel difference feature of every pixel to classify all the pixels into two parts: payload or non-payload. The feature is proved effective both theoretically and experimentally. Furthermore, the proposed method is compared with the maximum a posteriori estimator for payload location aimed at LSB-M. When the embedding rate is low, the method performs much better than the maximum a posteriori estimator.

Abstract:In the application of big data theory, there are many large scale multiple classification problems for the diversity and complexity of real world. However, the hyperplane updating of traditional multiple classification methods are not balanced. And the learning efficiency of them are low, and they are not efficient for the complex multiple classification data. To solve this problem, this paper presents an improved dynamical active multiple classification method (DYA). By combining the definitions of deadlock and activation with the active multiple classification process, the proposed method controls dynamically the status whether the sample is to be involved in the active learning process with the updating of classifier in it. Meanwhile, the active learnin g method with sub-bit counter and rotation learning approach is used to the balance learning and updating of classifier. The experiment results demonstrate that the proposed DYA method can improve both the learning efficiency and generalization performance.

Abstract:The research aimes at opinion sentence identification and element extraction in sentiment analysis in Chinese micro blogs. In the aspect of opinion sentence identification, the authors propose a new algorithm to compute the micro blog semantic sentiment orientation using sentiment words and emotional impact factors. In element extraction, subject term classification and the association rule are applied, accompanied with a series of pruning, sifting and delimiting rules to extract evaluative objects in micro blogs. Through mutual filtering of opinion sentence identification and element extraction, the recall rate is improved further. The released data of the sixth Chinese opinion analysis evaluation is adopted as experimental data. The results show that the methods perform well in opinion sentence recognition and element extraction. The precision ratio, recall rate, and F -value of opinion sentence identification are 95.62%,54.10% and 69.10%, respectively. The precision ratio, recall rate, and F-value of element extraction are22.07%, 12.66% and 16.09%, respectively.

Abstract:As the feature of a time series, Shapelet holds a good interpretability. Shapelet is widely applied recently in behavior reorganization, clustering and outlier detection, et al. However, time series data are synchronized and multi-sources in domains of electric power operation monitoring, medical image processing and streaming media, The relevance among time series are ignored if only extracting Shapelet from single source independently. Thus, to analyze the relevance of Shapelets from different sources, p-Shapelet is proposed as a new feature expressing time lags among Shapelets based on Shapelet. Specifically, for mining pair of Shapelets with most significant time lags from different classes, a parallel algorithm called the p-Shapelet miner is designed. It evaluates pair of Shapelets from different sources by information gain, and find the one(p-Shapelet) maximums information gain. The effectiveness and efficiency of the algorithm is verified by experiments using CMU motion capture datasets.

Abstract:The smoothed l0 norm algorithm in compressive sensing introduces smoothed functions to approximate the l0 norm. The problem of minimization of l0 norm can be transferred to a convex optimization problem of the smoothed functions, which could be used efficiently for compressive sensing reconstruction. Aiming at the choice of appropriate smoothed functions and improvement of the robustness of the algorithm, a new smoothed function sequence with gradient projection method has been proposed to solve the optimization problem in this paper. Singular value decomposition (SVD) method has been further proposed to improve the robustness of algorithm,then the accurate reconstruction of sparse signal is realized.Experimental results show that the proposed algorithm improve ignificantly in both the reconstruction accuracy and the peak value signal -to-noise ratio under the same test conditions.

Abstract:Human computer interaction based on hand gesture is one of the most popular natural interactive modes, which severely depends on the methods for real-time gesture recognition. Here, an effective hand feature extraction method is described, and the corresponding hand gesture recognition method is proposed. First, based on a simple tortoise model, one segments the human hand images by skin color features and tags on the wrist, and normalizes them to create the train set. Then feature vectors are computed by drawing concentric circles according to the center of the palm, and linear discriminant analysis (LDA) algorithm is used to deal with those vectors. Finally, an improved K-nearest neighbor (KNN) algorithm is presented to achieve real-time hand gesture classification and recognition. Experimental results with a self-defined hand gesture data set and multi-projector display systems prove the efficiency of the new approach.

Abstract:The region information of images is used by image segmentation based on C-V model, but features reflecting the detail of images such as edge information is ignored. For getting better results of image segmentation, it is particularly important to deal with these details. The amplitude of an image is larger in the edge region and smaller in the homogeneous region. It can be used to reflect the edge information of an image. By incorporating edge information into C-V model, using both the information of homogeneous regions and the edge information to control the active contours, it will obtaia better results of segmentation. The proposed model can overcome some disadvantages of C-V model, and achieve better image segmentation for those images that have the intensity inhomogeneity in backgrounds or weak edges.

Abstract:AdaBoost algorithm is a typical ensemble learning framework. It linearly combines a set of weak classifiers to construct a strong one, whose classification accuracy, generalization error and training error are all improved. However, the AdaBoost algo rithm is weak interpretability since it cannot simplify weak classifiers from output model. Hence, one presents a new algorithm, ensemble evolve classification algorithm for controlling the size of final model (ECSM), by introducing the genetic algorithm into the AdaBoost algorithm model. Gene evolution and fitness function can mandatory reserve the species diversity of samples in the AdaBoost iteration framework, and leave better classifiers. With keeping the classification accuracy, experimental results show that the proposed algorithm greatly reduce the number of classifiers compared with the classical AdaBoost algorithm.

Abstract:As an important part of the household appliances, the refrigerator becomes more intelligent. Object recognition of the food in a refrigerator is a key technology of a smart refrigerator. However, the foods in the refrigerator are diverse and disordered, which brings a lot of challenges to identify the varieties of foods. A method using an integrated convolutional neural network is proposed to solve this problem. The basic idea is that two convolutional neural networks are firstly trained separately. One is used to recognize the kinds of fruits and vegetables, the other is to recognize the color of them. Then, a multilayer perceptron is used to integrate the two independent networks to carry out classification. The two separate convolutional neural networks can complement and improve each other in the integrated network. In the method, color information, an important feature in the recognition, can be enhanced. The proposed structure also improves the recognition rate which is influenced by object occlusion and view variations. Finally, the effectiveness of the proposed method is validated on a dataset which contains a large amount of images obtained from a real situation of a refrigerator.

Abstract:In the knowledge base there exist characteristics of sparse for a single entity, and it is difficult to determine the similarity threshold of clustering. Therefore, this paper presents a name disambiguation algorithm based on cluster by step. Firstly, query features for character attribute are obtained from knowledge base, and the initial clustering based on knowledge base is carried out by text retrieval, which make up characteristics of sparse for a single entity name defined in knowledge base. Then, taking initial clustering results as input, name disambiguation in knowledge base is completed by using hierarchical clustering algorithm based on adaptive threshold. Finally, the other classes are identified by conditional random fields, and the cluster by using hierarchical clustering algorithm based on adaptive threshold is completed. The experiment on data of CLP2012 Chinese person name disambiguation results shows that the proposed algorithm can effectively achieve disambiguation names.

Abstract:Focusing on the data redundancy and load-unbalancing problems in the ex isting parallel FP-Growth algorithm, an improved algorithm for redundancy pruning and load balancing is proposed. Firstly, the improved algorithm introduces group task estimation method when using high-frequency strategies group. The longest path in the maximum pattern tree and the highest frequency are used as estimation. The group task will be averaged to others again when the group estimated is much larger than the value of other group. Then, the repetitive elements are removed in the list of different groups. Experimental results show that the improved algorithm avoides the data skew in the MapReduce and it is superior to the original one due to its high execution efficiency and speedup.