Abstract:Research on eye detection based on computer vision is very critical t o face recognition and gaze tracking. Furthermore, it is very important not only to related applications but also to the development of pattern recognition. Res earch findings on eye detection in recent 20 years are reviewed. Firstly, the existing classification methods are discussed and two new classification methods are proposed by recognition principle and detection results. Then, according to the detection results, the basic ideas, the advantages and the disadvantages of the eye detection methods are clarified. Meanwhile, the essence of all kinds of methods is tried to be revealed from different views and by different classification methods. Finally, the current situation of the eye detection research is analyzed and summarized, and the application prospect and development trend are discussed.

Abstract:Association analysis implemented with fantastic association measures is a basis of big data mining, so finding a reasonable measure is a key step for assocization analysis. Firstly, the challenge and research status of association measures are pointed out in the era of big data. From the perspective of the structure of the correlation measure, the exiting measures are systemized, and the properties and applicable corditions are summarized， respectively. Secondly, based on the development of correlation measures and the challanges of big data era, some conditions for meeting association measure are put forward to respond to meetting association measure challeges. Finally， some correlation measures in multi-modal data analysis are discussed and combed, and some ideas are provided to deal with the space conversion from three different angles, which attract more in depth thinking and research, therefore promoting the progress on big data mining.

Abstract:Least significant bit (LSB) matching algorithm and common steganographic methods, which use Gaussian support vector machine (GSVM) algorithm as the classifier, spend too much training time. Therefore, an improved logistic regression classifying algorithm named L curve truncated regularized iteratively re-weighted least squares(LTR IRLS) is proposed. Firstly, near optimal parameters of Tikhonov regularization are determined based on L curve, and convergence parameters of the truncated Newton algorithm are obtained through experiments for increasing the detection accuracy. Secondly, iteratively re-weighted least squares are utilized to search for the maximum loss expectancy and truncated Newton methods are utilized to avoid computing the Hessian matrix in the objective function, therefore reducing the computation amount greatly. Theoretical analysis and experimental results verify that LTR IRLS can ensure the detection accuracy rate higher than GSVM classifier, meanwhile reducing the training time and increaseing the detection speed.

Abstract:A method based on multi-threshold classification and attribute morphology is adopted to preprocess lunar image selectively in different gray layers, which ensure different highlight shadow crescent pair and low gray weak edge elliptical crater possess essentially normalized and steady Haar and PHOG feature. The influence and the function of the preprocessing method on partial wavelet Haar and pyramid histogram of oriented gradients feature is probed, and the effects of AdaBoost and SVM used in lunar crater detection are investigated. The integrated craters detecting strategy combining Haar and PHOG features with AdaBoost and SVM classifiers is also studied. The method is proved to have high accuracy and recognition efficiency. Experimental rusults demonstrate that lunar crater recognition radio is proved by 2%~5% via atlribute morphology and assemble classifier compared to traditional methods.

Abstract:Shuffle frog leaping algorithm(SLFA) combined with the grey correlation evaluation is proposed for designing orthogonal waveform. SLFA plays the dominant role in the new algorithm. Meanwile genetic operators of genetic algorithms(GAs) are introduced in the local position update and the group dividing method of SFLA is modified to improve the diversity of population. Moreover, the fitness function is evaluated with the grey correlation model. Polyphase orthogonal waveforms with low autocorrelation and cross correlation are taken as an example in the radar waveform design. Simulation results verify that the waveform generated by using the proposed algorithm obtains better orthogonal performance, thus the algorithm is effective for designing radar orthogonal signals.

Abstract:A bird identification method for the transient characteristics of birdsong signal based on adaptive optimal kernel(AOK) time frequency distribution identification is proposed. The collected birdsong signal is preprocessed and the spectrum is obtained through the AOK time frequency analysis method, Different energy distribution of birds sound signal at different time and different frequency are also analyzed. Then diagram spectrum is turned into gray image, the gray level co-occurrence matrix is calculated, image features is extracted as the eigenvalues of birds identification based on gray co-occurrence matrix parameters at different angles. Finally, the image texture of the known species is selected to generate training template and the image texture characteristic parameters of the species for identifying is used to generate the test template, Template matching is achieved using dynamic time warping (DTW) algorithm. The matching value are compared to find the minimum matching value corresponding templates, therefore the recognition of birds are realized. Finally, 40 kinds of common birds experiments demonstrate that the overall recognition rate reaches 96%.

Abstract:Considering traditional DS evidence theory deficiencies existing in dealing with conflict evidence, the information entropy attribute is put forward based on the similarity between each evidences, which fixed the evidence classification properties. Combining the similarity attribute between each evidences, the evidence set could be divided into high credibility evidence, general evidence and conflict evidence. The sorted evidence set is given different importance coefficients, and is modified to improve. After modifying, the general evidence and high conflict evidence are closed to the high credibility of evidence opinions. Finally, DS combination rule is utilized to synthesis for the modified evidence. It is difficult for obtaining data by multiple sensors to establish the basic probability distribution function for the evidence. For the problem, making full use of the ability that rough set theory can deal with incomplete information and knowledge, the decision information table is obtained via the attribute reduction of rough set. The function values are assigned for the basic probability by the decision information table. Combining the rough set and the improved D-S evidence theory, the real data sets are measured by all kinds of sensors. The experimental results show that the improved method can not only effectively solve the conflict problem, but also reduce the uncertainty of evidence.

Abstract:Aiming at solving the problem of recognition for the type of linear block code and the convolutional code, a method of the channel coding type recognition based on the run feature is proposed. Namely, the difference of the run feature of the linear block code and convolutional code is analyzed in theory. It shows that the randomness of the convolutional code run feature is superior to that of the linear block code, and the linear block code run number around the information bit length has a certain amount of distortion. Therefore, the code type can be identified by abstracting the run feature of these two codes. Simulation results show that the proposed method has high robustness and accurate recognition, and the method has a certain value in future engineering application.

Abstract:Community structure is one of the most important topological characteristics in the complex network, being a hot research area in different fields. A novel community detection algorithm is proposed based on edges rank and modularity optimization. Local graph is sparsificated and edges are ranked according to the similarity. Therefore, a method called the fast rank based community detection (FRCD) by maximizing modularity and fast mergement of edges is achieved. Meanwhile the method is also extended to dynamic and real time community detection on the basis of initial community structure, and a fast and robust dynamic community detection algorithm called the incremental dynamic community detection (IDCD) is presented. Theoretical analysis exhibit that FRCD has linear complexity for network edges. Experimental results in real world and artificial networks demonstrate the high accuracy and good erformance of the algorithm on static community detection and tracking dynamic structure of networks.

Abstract:Under the multi-sensor discrete system with correlated measurement noises and different unknown measurement functions, the distributed fusion Kalman filtering algorithm is proposed according to the existing fusion algorithms based on Bayes estimation weighted least squares (BYEWLS). The algorithm makes full use of the prior information for the unknown parameters. BYEWLS algorithm gains an advantage over WLS algorithm according to risk function. Meanwhile, the online method is proposed to eliminate the bias emerged with BYEWLS fusion algorithm. The distributed fusion algorithms can reduce the computational burden and improve the fusion accuracy, therefore they are suitable for real time applications. As a result, the simulation example indicates the validity of the theory analysis.

Abstract:In order to further improve the performance of speaker recognition system based on the GMM independent of text, a new speaker recognition method is applied to the speaker recognition system with small samples and text independent. Aiming at the large quantity demanded of training data during the modeling of the GMM, the advantages of the fuzzy set theory, vector quantization and the GMM are considered. Then through replacing the output probability function in the traditional GMM with the error scale of the fuzzy VQ, the requirements of the training data amount are reduced while improving the accuracy and recognition speed of the model. Meanwhile as a result of the fuzzy set theory playing a role of ＂plastic date＂, the similarity in the data of the target speakers is enhanced. Experimental results exhibit that the speaker recognition system of the method for the small sample data, achieves a superior recognition performance than the traditional speaker recognition system based on the GMM.

Abstract:A girth ACE union (GAU) construction algorithm is proposed for variably long LDPC codes considering the degree distribution, girth and approximate cycle extrinsic (ACE) message degree in a unified architecture. Therefore the short cycle number is reduced and the extrinsic message degree is increased. Large range of variable code length is accommodated through parity check matrix constructed by GAU algorithm. For short length codes, the performance of GAU algorithm is similar with that of LDPC codes in IEEE 802.16e. For middle length and long codes, GAU algorithm performs a little better. Variably long LDPC codes are designed using a common parity check matrix. The design of encoder and decoder with the same architecture for variable lengths is facilitated, which reduce the realizing complexity of encoder and decoder. Results show that the algorithm is useful in designing LDPC codes, which support variably long data communication, and the research is meaningful both in theory and practice.

Abstract:Since many methods for cluster user interest does not consider the semantic similarity of the user labels, a micro-blog user label interest clustering method is introduced based on feature mapping. Firstly, the user labels of the target users and their focus users are obtained, then the labels with the higher frequency than the threshold value is chosen. Therefore, a feature space is created. Secondly, the user labels are mapped to the feature space by calculating the semantic similarity based on the feature mapping. Finally, the fuzzy clustering is utilized to obtain the clustering result of different threshold value. Experimental results show that the method greatly improves the clustering accuracy rate for user interest clustering.

Abstract:Most research for anti-chaff using polarimetric information at home and abroad are considered without the influence of incident polarization both on scattering amplitude and scattering polarization. So a group of polarimetric scattering statistical data in given incident conditions is acquired and processed based on polarimetric coordinate warped, and preferential statistic curves of chaff cloud aiming at incident wave polarization form are also achieved through building a instantaneous polarized scattering mode of noncoherent chaff cloud in radar sight direction. Numerical simulation results show that the processed polarization states almost totally distribute on horizontal polarimetric area and chaff cloud has statistic preferential characteristics when incident polarization is at a minor angle of pitch of radar. Conclusions can be regarded as primary optimizing projects for anti-chaff using polarimetric information and provide the theoretical base for the polarization filtering.

Abstract:In the area of modern electronic warfare, two or more radiation sources are always placed within a certain range to prevent the attack from anti-radiation missile. These radiation sources are placed closely and their parameters are identical. In the case of multiple radiation sources, traditional direction finding algorithms of wideband signal cannot achieve good direction finding precision. Therefore, the single baseline system is set. New signals are created by the conjugate multiplication of signal spectrums received by two antenna array. According to the spectrum structure relationship of new signals, the time delay estimation of radiation sources can be obtained indirectly. And the precision of the time delay can be improved by wiping off the direct current component. The method has the feature of simple equipment. Simulation results show that the algorithm has good direction finding precision under certain conditions.

Abstract:The existing adaptive neighborhood graph embedding method based on local discriminant projection(LADP) only uses discriminant information in the principle space of local within class scatter matrix, which leads to the loss of discriminant information in the null space. To overcome the drawback of LADP, a complete LADP(CLADP) is proposed for face recognition. In the null space of local within class scatter matrix, irregular discriminant features are extracted by maximizing the local between class scatter matrix. In the principle space of local within class scatter matrix, regular discriminant features are extracted by maximizing the local between class scatter matrix and minimizing the local within class scatter matrix. Finally, irregular discriminant features and regular discriminant features are combined as the features of CLADP for face recognition. The experimental results on ORL, Yale face database and PIE subset illustrate the effectiveness of the proposed CLADP.

Abstract:Traditional fractional time delay estimation (TDE) algorithm is se nsitive to ambient noise and reverberation noise, and the algorithm performance is severely degraded in the complex environment. To improve the accuracy of time delay estimation, a GCC maximum likelihood phase compensation (GCC MLP) phase compensation fract ional time delay estimation approach is proposed based on generalized cross correlation(GCC). The frequency domain weighting function is also improved for GCC algorithm. At the same time, linear phase compensation is applied to the cross correlation spectrum in frequency domain and a continuous time delay estimator is obtained. Simulation comparition between the proposed algorithm and the other fractional TDE algorithms demonstrates that the proposed algorithm outperforms the curve fitting algorithm and the sinc interpolation algorithm. Meanwhile the robustness to the noisy and the reverberant conditions is reinforced, the time delay estimation error also decreases. Therefore, GCC-MLP phase compensation fractional time delay estimation algorithm can be widely used in sound source direction of arrival estimation and sound source localization.

Abstract:Blind source separation (BSS) algorithms based on the noise free model are not applicable when the SNR is low. To deal with this issue, one way is to denoise the mixtures corrupted by white Gaussian noise, firstly, and then utilize the BSS algorithms. Therefore, a Waveshrink algorithm is proposed based on translation invariant to denoise mixtures with strong noise. The high frequency coefficients sliding window method is utilized to estimate the noise variance accurately, and Bayesshrink algorithm is utilized for a more reasonable threshold. Consequently, the scope of the translation invariant is narrowed without degrading the performance of denoising, thus reducing the computation amount. Simulation results indicate that the proposed approach perform better in denoising compared with the traditional Waveshrink algorithm, and can remarkably enhance the separation performance of BSS algorithms, especially in the case with low signa.

Abstract:Device free localization (DFL) utilizing the received signal strength (RSS) variations on the wireless link caused by an object is the estimation of the object such as a person without carrying any electronic device. However, environment influences, such as temperature or swaying of sensor nodes, also alter RSS variations, thus degrading the positioning accuracy. A novel wavelet denoising algorithm based on subspace decomposition, combined with the fingerprint method, is proposed to reduce the environment impact. The noise characteristics of static environment are researched, then the maximum characteristic value is extracted as threshold, the signal features are adaptively decomposed into different orthogonal subspaces and the object signal is reconstructed in subspace. The feather extraction method is discussed after the mixed denoising analysis. Gaussian radial basis function is utilized to calculate the kernel distance between online measurement received signal strength and the fingerprint data to estimate the target location coordinate. Simulation results indicate that the proposed algorithm can achieve better positioning accuracy than the traditional location algorithm.

Abstract:To detect the phase quantized digital radio frequency memory (DRFM) deception jamming in realistic radar scenarios, an adaptive detector for detecting noise, jamming or echo signal in homogeneous environments is designed. Firstly noise or ″signal″ (filtered echo signal or jamming), is detected through adaptive matched filter (AMF)detector based on generalized likelihood ratio test (GLRT). Then the difference between two space time steering vectors of echo signal and jamming is used to redesign the detector to discriminate the echo signal or the jamming. The detector performance is assessed and analyzed through Monte Carlo simulation and theory deriation, which is compared with clairvoyant detector. Simulation results show that the detector can correctly detect the jamming signal under conditions of low quantization bits and high signal to noise

Abstract:Learning automaton （LA） is an adaptive decision maker that learns to choose the optimal action from a set of allowable actions through repeated interactions with a random environment. In most of the traditional LA, the action set is always taken to be finite. Hence, for continuous parameter learning problems, the action space needs to be discretized, and the accuracy of the solutions depends on the level of the discretization. A new continuous action set learning automaton （CALA）is proposed. The action set of the automaton is a variable interval, and actions are selected according to a uniform distribution over this interval. The end points of the interval are updated using the binary feedback signal from the environment. Simulation results with a multi-modal learning problem experiment demonstrate the superiority of the new algorithm over three existing CALA algorithms.

Abstract:Determination of the basic probability assignment (BPA) is the first and main step to the evidence theory application. How to generate BPA is still an open issue. To solve the problem, a method for determining BPA based on the cloud model is proposed. Firstly, the normal cloud model of each sample under the property is constructed based on the backward cloud generator. Secondly, through the antecedent cloud generator repeatedly, the average certainty of the test sample under this property is obtained. Thirdly, a method for measuring the similarity of normal cloud models is proposed, and the maximal similarity of the normal cloud model is made, which has the maximal certainty as the belief of the universal set. Finally, the certainty is normalized to obtain the BPA of each class. The effectiveness of the method is proved by experiments, and it can generate BPA in the case of little samples numbers.

Abstract:To solve the problem of high false alarm and low real time for aerial video with rotational ground background in moving object detection, a real time algorithm based on optical flow method is proposed. The image feature points are detected first, the optical flows are calculated, and the motion vector of the image background is estimated through optical flow field. Then the target size is calculated according to the flight height and the image is divided into blocks. Each feature point optical flow is compared with the motion vector of the image background to find out the candidates of the target feature point. Finally the candidates is added up in each block and the target areas are determined. Moving objects detection experiment is achieved for 360 pix×432 pix central area of two groups of experiment video. The result indicates that the algorithm can exactly detect the ground moving object with a low false alarm. The average consuming time is 29.460 ms/frame and 31.505 ms/frame, thus satisfying the real time requirements.

Abstract:The data in information system is dynamically changed.How to acquisite useful information according to dynamical varied information system is a key problem in data processing. To deal with the problem, the approaches for dynamical approximations acquisition while adding or deleting an attribute are respectively discussed in information system. By dividing original equivalent classes in information systems, an approach which avoids re-division of the universe is proposed. The efficiency of dynamical updating approximation is improved. By analyzing the relationship between equivalent classes and original approximations, the corresponding theorems between updated approximations and original approximations are given. Then, the approaches for dynamical acquisition of approximations while adding or deleting an attribute are respectively proposed in classical rough set model. Experimental results verify the validity of the approaches and prove that the efficiencies of the proposed approaches are better than those of the original approaches.