Abstract:Thresholding is one of the most widely used image segmentation methods because of its concision and effectivity. Many researchers have followed this technology with great interest and a lot of research findings have been published. Twenty years ago, the first author made a stage review on research progress of thresholding within 1962—1992. But at the present time, thresholding technology has made a big advance and new ideas have been proposed constantly. This paper aims at summarizing and classifying the commonly used hresholding methods in recent twenty years including brand new methods and improvements of the classical methods. Basic thoughts or criteria are given and characteristics and applicable conditions of these methods are clarified. It is hoped that this summarization can provide some thoughts and inspirations to the thresholding researches in future.

Abstract:The low-dose CT scans can effectively reduce the dose exposure of patients．　Ｈowever, the imaging quality will be lowered in the meantime. The way of maintaining high quality image that available for clinical diagnosis has become the main research direction in the field of CT technology. This paper summarizes the development and implementation of low-dose CT scans from the following aspects: low-dose scan implementation; statistical model of measurements; reconstruction methods and image processing methods. The research status of low-dose CT technology is also summarized. Finally, both the current research focus and the future research prospect are discussed and analyzed.

Abstract:With the increase of big data, solving Lasso problem becomes top research field. Past methods could not satisfy the time and efficient problem under big data situation. In order to deal with difficulty of computation and storage bringing from huge-scale and high-dimension data, this paper analyze the recent Lasso algorithm from three aspects: one-order method, random, and parallel and distributed computation, which play an important roles in dealing with huge-scale data problem. Based on those three aspects, this paper introduces and analyzes the novel algorithms: gradient descent method, Alternating Direction method of multipliers (ADMM), and coordinate descent method. Gradient descent method combine one order method and Nesterov's accelerate and smoothing method；ADMM put the random algorithm into the recent research; Coordinate descent make use of the character of coordinate system incorporation one-order method, random, and parallel and distributed computation. Moreover, this paper makes a deep analysis and research from primal and dual objective function.

Abstract:Spatio-temporal trajectory pattern mining has emerged as an active research field, focusing on the research and development of mining technology for big spatio-temporal trajectories to discover useful rules and knowledge. This paper attempts to review the recent research progress in spatio-temporal trajectory pattern mining and knowledge discovery. Then the background， application and advances of spatio-temporal trajectory pattern mining are introduced. And the research contents, system infrastructure and key technologies in big spatio-temporal trajectory pattern mining are discussed. Finally, the mining algorithm ideas for frequent pattern, flock pattern, gathering pattern, outlier pattern of spatio-temporal trajectory are expounded.

Abstract:The problems of efficient codec and resiliency to channel errors are important in video processing of wireless multimedia sensor networks (WMSNs). Based on the compressed sensing (CS) and dictionary learning algorithm, a dictionary learning-based compressed video sensing codec model is proposed for the WMSNs. The model uses CS to reduce the complexity of encoder effectively and improve resiliency to channel errors. In the encoder, the application of difference structure and skip mode reduces the amount of data transmitted in the channel. And in the decoder, dictionary learning algorithm helps enhance images′ sparse representation, thereby improve reconstructed video quality. The model switches the computational complexity from the encoder to the decoder and has high coding efficiency, so it can be applied to the recource-constrained embedded devices. The theory analysis and experiment results have verified the feasibility and efficiency of the model.

Abstract:Brain network aims to study the interaction of brain functional regions as a whole system, which plays a very important role for the understanding of brain function and structure, as well as the diagnosis of some brain diseases. As an important tool to analyze brain networks, machine learning has become a new focus of research since, and it can obtain the rules via automatically analyzing data and apply these rules to predict the unknown data. This paper reviews the concepts, methods and applications of brain network analysis, and mainly discusses some related works based on machine learning techniques from the following three aspects, i.e., construction of bran network, feature learning and classification and prediction. Finally, The conclution is drawed, and some new directions for future research is forecasted.

Abstract:In recent years, learning with weak supervision has become one of the hot research topics in machine learning. As one of the important weakly-supervised machine learning frameworks, partial label learning has been successfully applied to a number of real-world applications. In partial label learning, each object is described by a single instance (feature vector) in the input space. On the other hand, it is associated with a set of candidate labels among which only one is valid. The state-of-the-art on partial label learning researches is reviewed. Firstly, the problem definition on partial label learning as well as its differences and similarities with other related learning frameworks are given. Thenseveral representative partial label learning algorithms along with one of our recent progress on this topic are introduced. Finally, possible future investigations on partial label learning are briefly discussed.

Abstract:Image auto-annotation is a basic and challenge task in the image retrieval work. The traditional machine learning methods have btained a lot achievements in this field. The deep learning algorithm has achieved great success in image and text learning work since it is presented, so it can be an efficient method to solve the semantic gap problems. Image auto-annotation can be decomposed into two steps, that is, the basic image auto-annotation based on the relationship between image and tag, and the annotation enhanced based on the mutual information of the tags. In this article, the basic image auto-annotation is viewed as a multi-labelled problem. Therefore the prior knowledge of the tags can be used as the supervise information of the deep neural network. After obtained the image tag s, the dependent relationship of the tags is used to improve the annotation result. Finally, the model is tested in Corel and ESP datasets, and results prove that the method can efficiently solve the image auto-annotation problems.

Abstract:The existing similarity measurement methods for stock time series always ignore the trading volume and other important factors influencing the stock price. This phenomenon results in inaccuracy when clustering and classifying the series. To solve the problem, a new similarity measurement method for stock time series is proposed. The method which is based on dynamic time warping(DTW） introduces time-exhaustion factor and trading volume factor, and puts forward the ultimate similarity measurement formula for stock time series. To prove the feasibility and validity of the method, the stock time series in the household appliances and two others in the experiment of this paper are tested. The test result indicates that the new similarity measurement method based on DTW can maintain the shape features of stock series. On this basis, the method can solves the problem of price-volume relationship in the stock technical analysis well. Thus the method can be applied to pattern discovery and other fields in the stock technical analysis for more effective results.

Abstract:Lots of features in high-dimensional data are redundant or irrelevant. To tackle this problem, the concept of feature selection is introduced. In the meantime, many problems in machine learning involve examples that are naturally comprised of multiple views and with a limited number of labels. Multi-view learning and semi-supervised learning become the hotspots in machine learning. Hence authors investigate how to select relevant features with minimum redundancy from multi-view data with a limited number of labels, and propose a semi-supervised feature selection and clustering framework. To remove redundant and irrelevant features, authors exploit relations among views and relations among features in each view, and use a limited number of labeled data to help feature selection. The proposed framework in multi-view datasets is systematically evalated, and the results demonstrate the effectiveness and potential of the proposed method.

Abstract:As the social networks (such as Sina Weibo) become more and more popular and significant, spammer′s behavior severely affects the cr edibility and readability on social network platforms. A spammer detection model along with the algorithm is proposed based on duplicate detection of microblog posts to detect spammers on Weibo platform. Based on analyses of real-world data, the model is built by considering user behavior information, user social network information and content information. Experiments on a collection of real Weibo data shows the effectiveness of the proposed model. Parameters′ impaction to the model is also studied. The improvement of incorporating behavior information, content information and network information has been analyzed, hence the model is promising.

Abstract:A feature selection algorithm using quadratic programming is proposed based on feature margins. Firstly, the inner-class distance of features is taken as the coefficient of the quadratic terms in the objective function and the inter-class distance of features is used as the coefficient of the linear terms for searching informative features. The elements of the quadratic terms and the linear terms are normalized to balance the feature relation between inner class and inter-class. Then, the optimal solution vector is taken as the feature weight vector for selecting informative features. Finally, experiments on six different datasets show the effectiveness and feasibility of the proposed method.

Abstract:How to identify the semantic orientation of terms and build a high-quality sentiment dictionary to improve the accuracy of sentiment analysis on Micro-blogs has significant importance. Traditional algorithms based on corpus are sensitive to the seed words, and cannot effectively identify semantic orientation identification on low-frequency terms. To solve this problem, an algorithm based on word affinity measure is proposed to identify the semantic orientation of terms from Chinese Micro-blogs. Firstly, candidate words are extracted by the part of speech combination patterns. Secondly, Micro-blog emoticons are selected as seed words, and word affinity networks are built. Then, low frequency words are expanded by a synonyms dictionary during calculating the semantic orientation similarity between candidate words and seed words. Finally, the semantic orientation is determined according to the threshold. Experiments are conducted on a corpus with two million Micro-blogs using the proposed algorithm and traditional algorithms respectively. Experimental results show the advantage of the proposed algorithm.

Abstract:To apply supervised learning method in single face recognition problem, an improved algorithm based on sample augments by sliding window is proposed. The recognition time of the proposed algorithm is shorter than that of the original algorithm because of less feature dimension. Moreover, the mirror samples are generated to constitute auxiliary training set and two subspaces can be obtained by subspace learning. The recognition result is from the decision fusion of two subspaces and is robust to variation of the test samples. The experiment results on ORL face database and subset of Feret face database show that the proposed algorithm has higher recognition accuracy than other similar algorithms.

Abstract:Although much work has been done to elucidate the regulatory mechanism of miRNAs by associating miRNAs with mRNAs, their precise functions are still largely unknown. Latent dirichlet allocation (LDA) topic model is thus proposed to infer regulatory modules of miRNAs and their targets mRNAs for specific biological conditions. The proposed model firstly uses Welch′s t-test to mine differentially expressed miRNAs and mRNAs, and then a collapsed Gibbs sampling method is utilized to estimate parameters. The results on epithelial to mesenchymal transition (EMT) data sets show that the inferred functional miRNA mRNA regulatory modules (FMRMs) can construct regulatory relationships between miRNAs and mRNAs in different biological conditions, and give new insights into EMT biological process and miRNA targets therapy. Compared with K-means clustering algorithm, LDA topic model is more efficient in mining FMRMs.

Abstract:A region duplication image forgery detection algorithm based on Harris feature points and annular average representation is proposed. Firstly, an adaptive Wiener filter is applied to the image, and then Harris operator is utilized to extract feature points in the image. Secondly, a feature vector matrix is constructed with average values of pixels to make a quantity description of annular neig hborhood around each feature point, and lexicographical sorting and threshold processing are employed to implement similarity matching with the purpose of determining the candidate matching points. Finally, random sample consensus (RANSAC) algorithm is used to eliminate the erroneous matching points, and then the duplicated and tampered regions are located with identifiers. Experimental results show that the proposed algorithm is robust to rotation and flipping transformation of the copied region, and it can effectively resist common post-processing attacks such as Gaussian blurring, AWGN, JPEG compression and their mixed operations, especially the copy-move forgery with flat area of little visual structures and small area.

Abstract:Texture information of the single mode hands image is not sufficient for recognition in non-contact condition. A single device under the multi-spectrum is put forward to rapidly capture the images of palm print and palm vein. The texture information is enhanced and the recognition performance is improved through enhancing the palm print mainlines of the vein image. Firstly, the interval of acquiring the palm print and palm vein images are limited to less than 0.1 s, and the tiny displacement of palm is ignored. And the ROI of the palm print image is determined, whose coordinates are used to determine the ROI of palm vein image. Then, the preprocessing of the two ROI regions are performed, and the palm print mainlines are rapidly determined via the palm print high-frequence information from multi-layer wavelet decomposition. Finally, wavelet coefficients of palm print and palm vein are fused according to the high frequency component of palm print mainline normalized, and the blended texture image with the enhanced palm print mainline in the vein texture image is acquired. The experiment results show that the information of palm print mainline texture is significantly enhanced in the fusion images in the condition of keeping the original state of the vein texture, and the recognition performance is improved.

Abstract:To accomplish the automatic classification of thyroid nodules, the local texture features combining with the multiple instance learning method is proposed to overcome the overlap of the thyroid nodules. The local texture features are abstracted from the region of interest which is taken as the instance package composed of local features. The citation-kNN algorithm of the multi-instance learning(MIL) method is adopted to classify samples of this paper. Experimental results show that the identification method has higher classification accuracy and the accuracy achieves 85.59%. It is expected to be applied to the clinical diagnosis of thyroid, and provide a valuable reference for other related domains.

Abstract:Image segmentation is a key step in image analysis and image understanding. Compared with other image segmentation algorithms, mean shift algorithm has some advantages such as simple principle, dispensing with a priori knowledge, capability of dealing with gra y images and complex natural color images, etc. However, the algorithm requires iterative calculations for each pixel in the image, and segmentation computational cost is high for practical tasks. Therefore, a fast mean shift (FMS) method for image segmentation is proposed, in which a small amount of pixels are selected as an initial point for iterative calculation, and other pixels are mergered to the existing classes according to the distance between the pixel and the class centers. As a result, the proposed FMS method reduces the iteration numbers of mean shift algorithm, and boosts the segmentation efficiency. Experimental results show that the proposed FMS method can obtain good segmentation results and higher segmentation efficiency.

Abstract:Entity relation extraction is one of the most important researches in the field of information extraction. Previous researches focus on extracting various kinds of lexical or semantic features from the context where the related entities appeared, and one kind of classifiers (such as SVM) is used to extract the entity relation, but this kind of methods ignore the impact of the classifier performance on the entity relation extraction. Since SVM classifier has low accuracy for the test samples near the hyperplane, a method based on double-vote mechanism is designed for determining the fuzzy SVM samples. In the method, SVM classifier is used to classify the non-fuzzy samples directly; then, k-nearest neighbors (KNN) algorithm is applied to classify the fuzzy ones. The experiment on the data provided by SemEval-2010 evaluation task shows that the method can improve the performance of the entity relation extraction.

Abstract:A near-infrared face recognition system using DaVinci technology (OMAP3530) and embedded technique is proposed. The system uses 850 nm wavelength LEDs to provide proactive near-infrared light. Coordinating by OMAP3530 and EPM570 processors, it achieves visible and near-infrared images in real-time. The software design is based on the Codec Engine framework. Within the proposed system, ARM is responsible for image acquisition, user interface and database management, while DSP focuses on the core algorithms of image evaluation, face localization, feature extraction and matching. The system takes full advantage of the rich interface and powerful ima ge processing capability of OMAP3530 and is optimized by the C and assembly languages. Even when the light intensity changes, the system can still obtain accurate and fast recognition results.

Abstract:The mutual information model is introduced into the educational data association patterns mining. A new mining algorithm of the matrix-weighted positive and negative association patterns from educational data is presented based on mutual information model, and the related theorems and their proof are given. The algorithm overcomes the defects of the existing algorithms for weighted association patterns. It pays special attention to the various weights of the itemset in database, and also uses a new evaluation standard of positive and negative association patterns. Hence the positive and negative association patterns obtained from the educational data get closer to reality. Analysis on these patterns shows that, the potential educational and teaching rules, as well as educational development trend are discovered, providing a scientific basis for management, decision-making and teaching reform in education. Experiment results on real educational data demonstrate that the proposed algorithm is effective and reliable, with important potential value in the educational data processing and analyzing.

Abstract:In order to improve the accuracy of NOx emission concentration prediction of the coal-fired boiler and more accurately monitor the NOx pollution, this paper proposes a prediction method based on the least squares support vector machines (LSSVM) and the improved particle swarm optimization (PSO). According to LSSVM forecasting theory as well as the uncertainty of LSSVM parameter selection, an improved PSO algorithm to optimize the parameters of the model is used, a model of NOx emission characteristics is established, and the prediction results are compared with the results of other two methods simultaneously. Results indicate th at LSSVM is an effective modeling method which has higher fitting degree; the combination of improved PSO and LSSVM can improve the prediction accuracy and the generalization ability, and LSSVM is superior to the other two parameter optimization algorithms in the NOx emissions concentration forecast.