Abstract:Interactions between proteins and ligands are ubiquitous and indispensable in the process of life. These interactions play important roles in the biological molecular recognition and in the process of signal transmission. Identifying the binding residues of the protein-ligand interactions has important scientific significance for protein function research, drug design and screening. Biocomputing method has become an important method for the prediction of protein-ligand binding residues. This paper first describes the common definition of the binding residues of the protein-ligand interactions. Next, a systematic category of protein-ligand binding residue prediction is proposed, and several prediction methods in each category are described. This paper then demonstrates the related protein databases and the evaluation indexes, and experimentally compares and analyzes the performances of some representative prediction methods on the corresponding data sets. Finally, the future research directions of protein-ligand binding residue prediction are proposed, which provide some references for relevant researchers engaged in the field of protein-ligand binding residue prediction.

Abstract:A theoretical model of target detection is established for a general radar system. Specifically, we employ the thoughts and methodologies of Shannon's information theory to research the location information and amplitude-phase information of targets in a radar detection system. In the proposed model, we derive closed expressions of location information and amplitude-phase information in a single target detection assuming the target follows a uniform distribution in the observation range for a constant scattering coefficient and Rayleigh distributed scattering coefficient correspondingly. And in the high signal-to-noise(SNR) regime, we obtain an analytic expression and Cramér-Rao Bound(CRB) of location information for a constant scattering coefficient. Theoretical analysis shows that the location information of the target is linearly proportional to both time-bandwidth product(TBP) of radar detection system and the SNR in dB. The simulation results validate the theoretical analysis and give two important stages of radar target detection, namely, target acquisition and target tracking. The work of this paper is of great theoretical significance for the design of real radar detection systems.

Abstract:Recently, physical-layer network coding (PNC), a novel digital communication system, has aroused extensive interest of researchers. Similar to most other digital communication systems, the symbol timing plays an indispensable role in PNC. The existing research results, however, usually assume the symbol timing is fully known and rarely discuss the problem of symbol synchronization for PNC. In this paper, in view of the present situation, according to the maximum-likelihood estimation criterion we propose a new algorithm with a low oversampling factor (samples per symbol) based on the orthogonal training sequences. The proposed algorithm has the double advantage of a low sampling rate and high precision. It is shown by simulations that with a same oversampling factor, the proposed algorithm offers up to a 10-fold increase in the mean square error (MSE) performance over the conventional optimum sample (OS) algorithm for signal-noise ratio (SNR) of 5 dB.

Abstract:Distortion will exist in the range profile of the received radar stepped frequency signal of maneuvering target. To solve this problem, the integrated cubic phase function (ICPF) method is presented in this paper. First, the received signal is characterised as a three-order phase signal and performed with 1st-order differential operation. Next, ICPF of the processed back data is calculated to estimate and compensate the acceleration through dechirp operation. In addition, the estimation and compensation of the velocity of the residual signal are operated as well. Then the target's high resolution range profile can be obtained by IFFT. As the estimation of acceleration and velocity requires 1-D searching merely, the computational amount is quite small. Finally, the simulation is carried out in the situations of noiseless and low signal-to-noise ratio, and the test results prove the validity of this method in the paper.

Abstract:Tensor model-based parameters estimation is a trend for radar signal processing. However, the existing tensor-model based algorithms cannot achieve a good compromise between estimation accuracy and computational complexity. A three-way compressive sensing (TWCS) based algorithm is developed for angle estimation in multiple-input multiple-output radar. Exploiting the multidimensional structure inherent in the matched filtered data, a third-order tensor signal model is formulated. To lower the storage and computing complexity, the high-order singular value decomposition method is used to compressive the tensor data. The kernel tensor is linked to the trilinear model thus the compressed direction matrixes are obtained. Thereafter, the sparsity of the targets in the background is utilized and two overcomplete dictionaries are constructed for angle estimation with optimization methods. Taking advantage of the inherent multidimensional structure of the received data, the TWCS algorithm achieves better estimation accuracy than traditional subspace-based algorithms. In addition, the TWCS algorithm does not require further pairing of the estimated angles. Furthermore, it could achieve the doppler frequencies of the targets. Simulation results verify the effectiveness of the TWCS algorithm.

Abstract:The accurate matching of high and low resolution image blocks is the key of self-examples super resolution algorithm. In the process of blocks matching of high and low resolution images, considering the importance of texture image block structure, a similarity metric model based on constrained texture image patch is proposed in this paper. By using this exact matching model, the detail of super-resolution result image is further enriched, and the image quality is improved also. The new algorithm has the particular advantage of improving spatial resolution of image only using prior information of single low-resolution image itself. The experimental results show that the proposed algorithm has a better super-resolution visual effect compared with the bicubic interpolation algorithm and the local self-examples super-resolution algorithm, and it also has a good performance in the objective evaluation index.

Abstract:Image registration is a valuable technique for medical diagnosis and treatment. Due to the inferiority of image registration using maximum mutual information, a new hybrid method of multimodality medical image registration based on mutual information of spatial information is proposed. The new measure that combines mutual information, spatial information and feature characteristics, is proposed. Edge points are used as features and obtained from a morphology gradient detector. Feature characteristics like location, edge strength and orientation are taken into account to compute a joint probability distribution of corresponding edge points in two images. Mutual information based on this function is minimized to find the best alignment parameters. Finally, the translation parameters are calculated by using a gradient descent algorithm. The experimental results demonstrate the high validation precision and excellent accelerating capability of the algorithm.

Abstract:In order to improve the ability of image segmentation to grasp significant areas, an algorithm based on grid local watershed method and fuzzy C-means (FCM) is proposed by combing super pixel thoughts and watershed algorithm. The algorithm first partition an image into non-uniform grids according to the variance. For each grid, watershed algorithm is applied with the best gradient threshold to reduce the loss of local information in global watershed. In this way, the significant basins of each grid are extracted. Through regional integration and mean normalization of each area, FCM clustering considering the size of each region is used to get the final segmentation image. Experimental results show its great robustness against noise. In addition, the algorithm can effectively eliminate the interference regions and segment the significant regions of images with a low time complexity.

Abstract:A multi view face detection algorithm based on multi-channel map discriminant projection HAAR feature is proposed. Firstly, the multi-channel map is extracted from the face image by the algorithm, which can reduce the influence of illumination and noise in the image. Secondly, based on the positive and negative training samples, the enhanced HAAR feature is learned by the linear discriminant projection, which can improve the distinguishing ability of the feature. Then the response in multi-channel of the augmented HAAR feature in the training sample is calculated, and the non symmetric GentleBoost algorithm is used to generate a set of weak classifiers. Finally, the weight and threshold of the strong classifier are adjusted by the linear non symmetric classifier. This method not only improves the distinguishing ability of the feature, but also realizes the reasonable division of the non balanced positive and negative sample space.Experimental results show that the proposed method has a faster detection speed and the higher detection accuracy compared with the classical methods.

Abstract:Automatic aortic image segmentation plays an important role in early aortic disease diagnosis, risk evaluation and treatment planning. In this paper, we use a multi-atlas based medical image segmentation method and first combine it with a joint label fusion(JLF) strategy to segment 3D aortic CT images automatically. Joint label fusion strategy takes the correlation of atlases into consideration and the effect of redundant information of atlases can be restrained. To handle the problem of insufficient atlases, we propose an atlas archive update method which can enhance the segmentation accuracy with relatively low computational complexity. We evaluate our method by using a data set with 15 aortic subjects and comparing with three widely used label fusion techniques (majority voting, local-weighted label fusion and STAPLE). Experimental results show superior performances of our method to state-of-the-art.

Abstract:Label propagation is a widely used community detection method with low complexity. It assigns an initial label for each node in the network, and then propagates the labels to discover the potential community structure in complex networks. However traditional label propagation is faced with some inadequacies, such as ignoring the difference between nodes and input parameters demanding. To overcome those defects, this paper puts forward an overlapping community detection algorithm based on the improved multi-label propagation. It uses K-shell decomposition method to identify core nodes of the network firstly, and then updates labels outward layer by layer. The number of labels of overlapping nodes is determined by the types of neighbor node when choosing label for a node. Experiment results show that this algorithm makes the community detection results more accurate and stable.

Abstract:Optical coherence tomography (OCT) plays a very important role in retinal/choroidal examination. However, image blurring can be introduced into the OCT images during the image acquirement process, resulted from the movement of eye or the out-of-focus effect of the imaging machine. Therefore, OCT image deblurring is of great importance in practice. In this paper, we use an expectation-maximization (EM) based image deconvolution technique for eliminating the blur effect in the OCT images. The proposed technique is designed based on several of the special characteristics of the OCT retinal image and is validated to be able to outperform various state-of-the-art image deblurring techniques as shown by the experimental results in the paper.

Abstract:Short time Fourier transform and wavelet transform are not effective in extracting features of electrocardiogram (ECG) signal for arrhythmia detection. Therefore,a novel algorithm based on the feature selection of S-transform is proposed for arrhythmia classification. First, ECG signals are processed by S-transform, and the time-frequency features are extracted from both the amplitude and the phase of ST results. Then, time-frequency features, morphological features, and RR interval are combined as the original feature vector. Second, the genetic algorithm (GA) and support vector machine (SVM) are combined as a Wrapper approach to search an optimal feature subset. The feature weights are computed by ReliefF algorithm, and the initialization of genetic population depends on the feature weights. Moreover,GA searches an optimal feature subset using classification performance as the fitness function. Finally, a multi-SVM model with one against all (OAA) strategy is built for the classification of eight types of ECG beats from the MIT-BIH arrhythmia database. Experimental results indicate that the proposed approach has the best performance among other state-of-the-art approaches, and the sensitivity, specificity, and accuracy reach 96.14%, 99.75%, and 99.81%, respectively.

Abstract:Feature selection is an important preprocessing step in machine learning and data mining. Feature selection of class-imbalanced dataset is a hot topic of machine learning and pattern recognition. Most traditional feature selection classification algorithms pursue high precision, and assume that the data have no misclassification costs or have the same costs. However, in real applications, different misclassifications always tend to produce different misclassification costs. To get the feature subset with minimum misclassification cost, a supervised cost-sensitive feature selection algorithm based on sample neighborhood preserving is proposed, whose main idea is to introduce the sample neighborhood into the cost-sensitive feature selection framework. The experimental results on eight real-life data sets demonstrate the superiority of the proposed algorithm.

Abstract:Aiming at singular value of matrix decomposition and ill-posed problems in traditional batch processing principal component analysis (PCA) based operational modal analysis (OMA), an operational modal identification method based on self-iterative principal component extraction is proposed. Different from traditional batch processing PCA, which obtains all principal components by matrix decomposition one time, the proposed method can realize the identification of main contribution operational modals by self-iterative principal component extraction one by one. Theoretical analysis shows its lower time and spatial complexity than traditional batch processing PCA based OMA. The simulation results on simple beam datasets show that the self-iterative principal component extraction algorithm can identify effectively main contribution modals and natural frequency of linear time invariant structure from smooth and random response signals. And it has smaller time cost in the case of more response points and more sampling time in contrast with traditional methods.

Abstract:Traditional user attribute inference method is mainly based on machine learning and statistical learning methods, and its inference method ignores the user's overall representation and the correlation between tasks. A user attribute inference method based on multitasking ensemble model is proposed, which uses doc2vec unique structural characteristics and adds document vector to achieve the overall representation of the user, thus avoiding the limitations of artificial features extraction. In order to realize the multi-attribute inference task, a multi-task ensemble framework based on association learning is proposed, which is to identify multiple attributes of a user individually and give the multi-attribute representation of a single user. It enhances the overall representation of user. The relationship between multiple attributes is established at the same time, so as to improve the distinguishing degree of single-task learning. Then, this paper uses the model ensemble technology to complete the inter-attribute learning, improves the accuracy of learning and model generalization ability, and uses as few models as possible to improve the model operation efficiency. Experimental comparison on several data sets shows some advantages over other algorithms.

Abstract:Convolutional neural network (CNN) is a kind of common architecture of deep learning, which is inspired by the biological visual cognition mechanism. CNN can obtain the effective feature expression from the original image. In recent years, CNN has made breakthroughs in the field of image recognition, but it takes a lot of time in the training process. As a new machine learning algorithm proposed by Leo Breiman in 2001, random forest (RF) has high accuracy in classification and regression, fast training speed and is not prone to over-fitting. The existing RF based classifiers rely on hand-selected features. Aiming at the above problems, a new C-RF model based on CNN is proposed in this paper, which puts the features extracted by CNN into RF to complete the classification.Since the network using random weights can also obtain effective results, gradient algorithm is not used to adjust the network parameters for avoiding a lot of time consumption. Experimental results on the MNIST and the Rotated MNIST datasets show that the classification accuracy of C-RF model is better than that of RF, and the generalization ability is also improved at the same time.

Abstract:An improved algorithm is proposed to correct the assignments of fuzzy points for the previous fuzzy collaborative clustering. The new expression of similarity distance based on clear radius is introduced, and the hypersphere central region is used to represent one cluster instead of the traditional center. In the light of initial results of separated subsets, the membership degrees are recalculated for the fuzzy points in which wrong assignments easily occurred, and finally the more clear-cut partition is obtained. The experimental results show that the improved algorithm can reduce the fuzzy points widely distributed near the boundary and correct quite part of the wrong partitions. Moreover, the method of clear radius can simplify the parameter setting by weakening the difference of collaborative coefficients.

Abstract:In multi-instance learning, the core instances play an important role on the prediction of bags' label. And if two instances have different numbers of instances with the same category around them, they have different levels of representative. In order to improve the classification accuracy, multi-instance learning with instance-level covering algorithm (MILICA) is proposed by which we could select the most representative instances to form the core instance set. Firstly, with the max Hausdorffdistance and the covering algorithm, the initial core instance set is constructed. Then, the final core instance set and the number of instances in a cover are obtained. Finally, a similarity measure function is used to convert a bag into a single sample for classification. Experimental results on two-category datasets and multi-category image datasets demonstrate that the proposed MILICA method has perfect classification capability.

Abstract:Traditional methods of traffic accident detection and review are mainly through manual monitoring. To overcome the low efficiency and poor performance of those methods, we proposed a vehicle abnormal event detection method based on the up-to-date video coding standard HEVC(High-efficiency video coding). Firstly, we extract the motion vector information from HEVC bit-stream to proceed cumulative iteration and median filtering. After that, we calculate the motion intensity value of the moving objects based on block partition and motion vector information, and then extract the moving targets according to the motion intensity value and eight-neighbored region method. Finally, we detect the vehicle abnormal events from the video sequence using the space distance and motion intensity evaluation method. Experimental results show that the proposed method can accurately detect the abnormal events in video sequence, especially for those videos with fast-moving objects and multiple moving targets.

Abstract:Image marking of the region of interest is an important research topic in image processing in recent years. The combination of low and middle levels can ensure the result has both of their information. First, we get the middle-level coarse saliency map by using the boosting Harris to make a convex hall and superpixels clustered by GBR. And then we weight different Gaussian filters to get the low-level saliency map. The final saliency map is combinated by middle-level saliency map and low-level saliency map. Experiments on the public databases coming from Microsoft Research Asia show that the proposed algorithm performs better than state-of-art algorithms not only on subjective evaluation but also on objective evaluation, and it is effective at the eliminate of background noise and outstanding at making the saliency regions high light.