Abstract:Multiple input multiple output(MIMO) radar is a new kind of system radar which introduces the multiple-input multiple-output technique from wireless communication field into the field of modern radars with the combination of digital array signal processing. Due to the waveform diversity, MIMO radar has many advantages compared with traditional phased array radar. To fully understand the MIMO radar, we make a comprehensive review for MIMO radar. Firstly, the concept and principles of MIMO radar are introduced, while the relation between the MIMO radar and phased array radar is revealed simultaneously. Then, based on the performance analysis of MIMO radar, we give advantages and disadvantages of MIMO radar in applications compared with conventional phased array radar. Finally, combined with the coherent characters and advantages of MIMO radar, several potential applications are introduced.

Abstract:Nerve cells image segmentation is of great significance for neuroscience research, but the complexity of the nerve cells submicroscopic structure and the quality problem of the missing and fuzzy of the boundary produced by transmission electron microscope(TEM) have making it being a problem in medical image processing. Based on the significant local clustering characteristics of nerve cell TEM images, a local feature-constraint information based TEM image segmentation algorithm is proposed by using the superpixels technology. First, superpixel structure is built based on the graph-based model. Then the local spatial information of superpixels based on MRF spatial neighborhood are extracted to solve the complex neighborhood information and space structure. Finally, the segmentation results can be obtained by the MRF model optimization and superpixels merging. The research results show that the proposed algorithm is accurate and robust with better describtation of submicroscopic structure.

Abstract:Since ‘particle impoverishment’ of the particle filter deteriorates the performance of the single channel blind signal separation based on traditional particle filtering and its huge calculation, a new single channel blind BPSK signal separation algorithm based on particle flow filtering is proposed. Firstly, according to single channel signal of mixing two BPSK signals, a measuring equation and a state equation are built. Secondly, particles are updated through moving the particles in state space which obey the prior distribution to its corresponding posterior distribution, which is different from using resampling to update particles in particle filtering (PF), thereby avoiding the ‘particle impoverishment’ phenomenon caused by re-sampling in PF. Lastly a weak solution based particle flow filter is used to achieve BPSK blind signal separation. Computer simulation results show that compared with particle filtering algorithm the new algorithm has the lower bit error rate and computational complexity.

Abstract:Based on MapReduce and upper sampling, an approach for imbalanced big data classification is proposed in this paper. The proposed method includes five steps:(1) For each positive instance, its nearest neighbor is found by MapReduce. (2) Some positive instances on the line between the two points are created. (3) According to the cardinality of the set of positive instances, the set of negative instances is partitioned into some subsets. (4) Some balanced subsets are generated with the set of positive instances and the subset of negative instances. (5) Some classifiers are trained by extreme learning machine on the generated balanced subsets, and the trained classifiers are integrated by majority voting for classifying new instances. Experimental comparisons with three related methods are conducted on five imbalanced big data sets. The experimental results show that the proposed method outperforms the three methods.

Abstract:Informative gene selection is an essential step to perform tumor classification with large scale gene expression profiles. However, it is difficult to select informative genes related to tumor from gene expression profiles because of its characteristics such as high dimensionality and relatively small samples, many noises, and some of the genes are superfluous and irrelevant. To deal with the challenging problem of finding an informative gene subset with the least number of genes but the highest classification performance, a novel hybrid gene selection algorithm named SUNRS is proposed based on the symmetric uncertainty (SU) and neighborhood rough set (NRS). Firstly, the symmetric uncertain index, which aims to eliminate redundant and irrelevant genes, is used to select top-ranked genes as the candidate gene subset. Secondly, the neighborhood rough set reduction algorithm is used to obtain the target gene subset by optimizing the candidate gene subset. Experimental results show that the proposed algorithm can obtain higher classification accuracy with less informative gene, which not only improves the generalization performance of the algorithm, but also enhances the time efficiency.

Abstract:Atmospheric turbulence can cause the image degraded with time-varying blur and geometric distortion. We resolve the object detection problem by proposing a three-step approach. According to the low-rank decomposition, the first step decomposes the turbulence sequence into two components:the low-rank stabilized background and the sparse errors. The sparse part in the result of first step includes turbulence distortion, noise and moving object. Then, the sparse matrix is processed with adaptive threshold to select the block-sparse mask and the holes within the mask are simultaneously filled. The low-rank matrix is processed with different Gaussian models to extract the foreground. Finally, a decision fusion module is introduced to exploit complementary information from two approaches to boost overall detection accuracy. The experimental results have shown the promising performances. Compared with traditional methods, the proposed approach can not only improve the detection rate, but also handle the interferences of strong turbulence.

Abstract:As the linear discriminant analysis (LDA) is just a linear method and is difficult to effectively deal with nonlinear problems, non-linearizing LDA is a crucial strategy to enable it to solve such nonlinear problems. Nonlinear LDA is mainly based on two strategies, neural networks and kernelization. A representative of the former strategy is the neural network discriminant analysis (NNDA). Athough NNDA inherits the advantages such as self-adaption, parallel processing, distributed storing and nonlinear mapping of neural networks, its training is quite time-consuming and likely to get trapped in local minimum. While the representative of the latter strategy is the kernel linear discriminant analysis (KLDA). Although KLDA can obtain a global optimal analytical solution, its computational cost is rather high, due to the fact that the number of hidden nodes of KLDA is equal to the size of training samples, especially in large scale scenarios. Inspired by the idea of random map, a novel extreme nonlinear discriminant analysis (ENDA) is proposed by reconstructing NNDA via extreme learning strategy in this paper. ENDA shares both the self-adaption of NNDA and the efficient computation of global optimal solution of KLDA. Finally, experimental results on UCI datasets demonstrate the superiority of ENDA over KLDA and NNDA in classification accuracy.

Abstract:The propagating ultrasound is disturbed by the clutter, which is resulted from interior discontinuities, material impurities, etc.. And the clutter may submerge the detected flaw reflection.Therefore, ultrasonic echo needs to be enhanced.Wiener filter is the common method of acoustic enhancement, and here is utilized in the ultrasonic noise reduction.We give the procedure of the Wiener filter to process the ultrasound.Firstly, initial noise power spectra of no pulse phases are statistically averaged.Then the power spectra of the ultrasound phase with clutter are adaptively calculated. Finally, the Wiener filter is performed to each ultrasonic phase. The experiments of the ultrasonic data filtering under different signal-noise-ratios are demonstrated that the proposed method is valid.Wiener filter has the same ability with spectral subtraction to reduce the noise superposed on the ultrasonic echo, and is more effective.

Abstract:As one of the hot issues of remote sensing imaging, the traditional method of remote sensing image compression has problems that widespread a long reconstruction time, and the quality of the reconstructed image needs to be improved. According to the remote sensing images of different typical surface feature, the K-SVD dictionary learning method is utilized in the paper. In the process of reconstruction, through multiple iterations on the part of the image blocks, the original image can be solved by a linear representation of the atoms from the corresponding surface feature of an overcomplete dictionary. Then the atoms are given preferentially as the initial value to calculate the residual of the image blocks in the neighborhood, to reduce the number of iterations. The remote sensing image information content on typical surface and the similarity between image blocks are fully exploited. Compared with the general dictionary structured by non-redundant orthogonal base or non-classified learning dictionary, the proposed method outperforms in the reconstructed image quality and reconstruction speed.

Abstract:The paper studied the multipath interference mitigation in global navigation satellite system (GNSS). A novel efficient weighted RELAXation (WRELAX) based algorithm considering the characteristic of the correlation of GNSS signal is proposed. The proposed algorithm has the advantage of lower computation burden compared with the state of art methods. In the proposed algorithm, by windowing the auto-correlation of the reference signal and the cross-correlation between the received data and the local reference signal, the data length in the iterative parameter estimation process is greatly shortened. Therefore, the amount of computation burden is reduced. Experimental results indicate that the proposed algorithm has similar performance and much low computation burden compared with the original WRELAX algorithm.

Abstract:The weight of antecedent attributes can't work accurately in the linear combinational belief rule based system usually. Simultaneously, with an increase in the number of evaluation ranks, the new weight activation formula will have negative effects on results. Aiming at the above drawbacks, this paper proposes a two-value and multi-base reasoning method based on the existing belief rule based inference classification algorithm to improve the belief rule based decision system. The evaluation of belief rules in the conclusion are divided into two ranks firstly, which means making a two-value judgment on a decision problem. Then many belief rule bases are set to solve some sub problems simultaneously. Finally results of many sub problems by multi-base reasoning method are mixed to solve the classification problem. Experimental results show the feasibility and effectiveness of the proposed belief rule base reasoning classfication method.

Abstract:Gathering pattern is an important research topic in the field of trajectory pattern mining. It focus on collective gathering problem on consecutive time period. Traditional models of gathering patterns are based on co-concurrence patterns. Mining methods based on such models generate a lot of stationary gathering groups. In order to deal with such problems, we propose a converging pattern based on modelling of group moving objects, which accurately identifies gathering group instead of other types of moving group. A moving objects converging pattern mining (CPM) algorithm is presented and implemented. First, the algorithm locates all high density peak points and converges central zones. Second, the algorithm identifies converging groups on consecutive timestamps, and then detects converging patterns according to the durability of group patterns. Experimental results show the effectiveness and efficiency of the algorithm.

Abstract:As a frequently personalized recommendation algorithm of the currently recommendation system, collaborative filtering uses the item evaluation by the approximate users to recommend. Kernel function is an approach for non-linear pattern analysis problems. Ordinarily, collaborative filtering will choose some different kernel functions to analyse the influence between the users. Since the single kernel function can not be adapted to the complicated and various scene, the combination of multiply kernel function becomes a solution. In terms of scenes, multiply kernel learning can combine every kernel function for a better result. This paper proposes a collaborative filtering algorithm based on multiple kernel learning. Based on the available kernel function, this algorithm optimizes the weights of every kernel function to match the data distribution. The experimental result on dianping dataset and foursquare dataset shows that compared with the collaborative filtering algorithm based on common similarity, the collaborative filtering algorithm based on multiple kernel learning achieves better performance. That is, multiple kernel learning has a better common adaptation.

Abstract:As a common dimensionality reduction method, the supervised Laplacian discriminant analysis (SLDA) for small size sample achieves a good result of dimensionality reduction via graph embedding discriminant neighborhood analysis. However, when SLDA finds the inter-class and intra-class data points in K nearest neighbors, there might exist an imbalance problem. Additionally, SLDA does not fully consider the inter-class information, which may decrease the performance of SLDA to a certain extent. To address the two problems mentioned above, we propose a double adjacent graph-based discriminant analysis (DAG-DA) algorithm for small size sample. Firstly, the algorithm tries to find K nearest neighbors in inter-class and intra-class samples, respectively, and then uses these K inter-class neighbors and K intra-class neighbors to construct the double adjacent graph. In this way, we can ensure that the adjacent graph contains both the inter-class and intra-class data points and has the same number. Secondly, the algorithm tries to add the intra-class Laplacian scatter matrix into the objective function of SLDA. Thus, the projection matrix obtained by optimization takes the information between classes into account fully. We perform experiments on Yale and ORL human face datasets. Experimental results show that the proposed algorithm can get better performance compared with other methods.

Abstract:Fire prevention has gradually dominated by video surveillance. There are many various algorithms of video flame, but most of them require lots of video flame samples to do the training for the final classification. The results often fail to achieve a high detection rate if flame samples are not enough. So a new method that uses the super pixel segmentation and the flash frequency for recognition is proposed in this paper. In the Lab color space, the cencroids of each region are represented as a super pixel after some approximate homogeneous regions are segmented in the flame picture based on super pixel segmentation. The candidate regions are extracted based on the RGB and Lab color features according to some rules. At last whether the candidate region is a flame is determined by the flash frequency characteristics. Experimental result presents good performance with high detection rate in the case of small samples.

Abstract:Pedestrian detection is a highspot and challenge research work in the area of computer vision and pattern recognition. The aggregate channel feature (ACF) algorithm generates lower detecting precision and higher log-average miss rate(LAMR) for pedestrian detection. We proposed an improve pedestrian detection method based on ACF algorithm in this paper. Firstly, we introduce objectness method to further verify low detection score object area captured by ACF, which can reduce false positive (FP) of the algorithm to some degree. Then, we combine the score with location of the detection window to modify the non-maximum suppression (Nms) algorithm, and the AP increases by 0.41%, while the LAMR decreases by 1.49%. Finally, we implement cascading detection for detection area by using a given threshold score and a casDPM model. The AP increases by 0.65%, and the LAMR decreases by 2.06%. Experiments on INRIA dataset are conducted and validated, and the results show that our approach not only meets the needs of real-time detection, but also obviously decreases FP, and displays a good detection effect.

Abstract:The purpose of cross-language topic discovery is to classify news texts written in different languages by their topics automatically. However, due to the difference in different languages, it's hard to describe these texts on the same feature space, so mining the same topic is not an easy work. When a particular news event is reported, the news elements are the same no matter which language describe it. So news elements can reflect the relevance among different news texts. Therefore, the paper proposed Chinese-Vietnamese bilingual news topic detection methods based on graph clustering. Firstly, Chinese-Vietnamese bilingual news elements are extracted and the similarity of different news texts is calculated by using the news elements' similarity to set up a Chinese-Vietnamese bilingual news graph model. Secondly, through the propagation characteristics of the Chinese-Vietnamese bilingual news graph model, the similarity matrix is adjusted by using the random walk algorithm. Finally, affinity propagation algorithm is used to cluster topic. The experimental result shows that the proposed method is effective.

Abstract:The existing group classification methods ignore the functional characteristics and their access probabilities implied in geographical positions. To solve this problem, a group classification method based on location semantic and probability is proposed, which includes two parts,location semantic discovery and the access probability vector clustering. Firstly, the location semantic implied in location words is obtained by using location semantic discovery method. Then according to the location semantic distribution, the access probability vector of mobile users for the location semantic space can be obtained. Finally, the group classification can be realized by using the access probability vector as the clustering weight vector. Experimental results show that the proposed method can effectively extract the location semantic coinciding with the reality and obtain similar users with similar access probabilities in location semantic space. Compared with the available group classification methods, the proposed method can achieve better experimental effects with an increase in F-measure of 4%.

Abstract:In this paper, a coarse-to-fine traffic sign recognition algorithm is proposed to alleviate the conflict between recognition precision and time consumption. In the coarse classification, a traffic sign region is represented with color name-histogram of gradient (CN-HOG) descriptors to describe its color and shape features. A linear support vector machine (SVM) classifier is used to classify the region into different categories:prohibitory, warning, mandatory, release of prohibitory and others. In the fine classification, the different fusion methods of color and shape features in Bag of Words model are discussed and the color-shape early fusion method is employed to combine the CN and scale-invariant feature transform (SIFT) descriptors. The final class labels of the region are obtained by Gaussian kernel SVM classifier. Experiments in public dataset show that the proposed algorithm satisfies real-time practice and meanwhile achieves a high classification precision of 99.15%.

Abstract:Compared with ensemble learning, the ensemble pruning is used to search for the optimal subset among multiple classifiers to improve the generalization performance of the classifier and simplify the ensemble process. In order to improve generalization performance and simplify ensemble process, ensemble pruning is used to search an optimal subset in multiple classifiers. It has attracted widespread concern, and it is significant to reduce the complex of ensemble learning. In recent years, researchers have proposed Pareto ensemble pruning (PEP) which considers both the classification performance and the number of base learners, and solves the two goals as the bi-objective optimization. However, Pareto ensemble pruning method ignores the diversity among classifiers, which would cause relatively large similarity among classifiers. In the paper, we proposed Pareto ensemble pruning with diversity (PEPD), in which diversity among classifiers is introduced into Pareto ensemble pruning method. The first goal of the proposed method is to maximize classifiers' diversity and their classification performance. The second goal is to minimize the number of base learners. The experimental results show that the PEPD method can obtain higher performance in most cases. And the enhancement is due to diversity's combination when PEPD and PEP have the similarity number of base learners. Experiments show that the PEPD method can obtain higher performance in most cases due to diversity's combination, when PEPD and PEP have the similar number of base learners.

Abstract:Service perception analysis is the crucial solution to understand user experience and network quality, and the maintenance and optimization of the mobile networks. With the fast development of mobile Internet and OTT(over-the-top) services, the traditional network-centric mode of network operation and maintenance is no longer an effective way. In this situation, how to effectively evaluate and optimize user's service perception is getting more and more important. The key factors that impact on user's end-to-end OTT service perception is analyzed based on the so-called crowd sourcing based user perception monitoring methodology. Furthermore, the inner relationship among the key factors and the interaction between key quality indicators are evaluated from several aspects, by utilizing big data set of user perception collected from large amount of smart phone users in the real network. The results of analysis are highly valuable references for the improvement of user perception and network quality, network deployment and optimizations, etc.