Abstract:Speech interaction technology is becoming increasingly popular in practical voice-driven applications. However, due to the inferences caused by reverberation in real-world environments, the performances of speech interaction in the distant-talking condition are far from being satisfactory. Decades of efforts are devoted to solving the reverberation problem and spawning a vast variety of practical methods. Recently, the deep learning technique, which is developing rapidly and has greatly reshaped the speech processing community, also acquires remarkable performance in speech dereverberation. However, a systematic analysis and summary of the inherent relationship between the recent deep learning based methods and the previous classical methods is rarely seen. As such, we give a comprehensive overview of the current and past development of single channel speech dereverberation. Then, the main challenges are discussed. Finally, we share some views of its future development.

Abstract:Aiming at the low convergence rate and high mean square error of traditional constant modulus algorithm(CMA) and too many parameters and high complexity of traditional neural network, a complex neural polynomial blind equalization algorithm based on nonlinear Volterra channel is studied. In the studied algorithm, the complex-valued neural polynomial with a single layer neural network and nonlinear processor has very simple structure and low complexity. And the fuzzy rule controller based on fuzzy neural network (FNN) can effectively control the step-size of scale factor. The simulation results show that the proposed algorithm not only has simple structure, low complexity, fast convergence speed and small steady-state error, but also can solve the contradiction between convergence speed and mean square error.

Abstract:A type of wide-band space-time constant modulus array algorithm for multi-user signal acquisition and interference suppression in underwater acoustic communication system is proposed to realize multi-user signal extraction and wide-band interference suppression. Compared with the conventional constant modulus array, it has stronger ability of multi-path and interference suppression. As an important part of cognitive underwater communication system, the time-frequency analysis and the instantaneous autocorrelation modulation algorithm for feature extraction provide an effective technique to classify the multi-user signals and interferences obtained by the wide-band constant modulus array, providing a solution to demodulation for different users and interference suppression. The performance of the cognitive processing structure is proved via numerical simulations.

Abstract:In the study of text detection, the proposal extraction method is not widely concerned and deeply studied, due to the structure of the text and otherness of the general object and the high precision requirement of text detection. In this paper, we propose a proposal extraction method for text detection. The proposed method firstly utilize the fully convolutional network to predict the text regions, which can effectively reduce the search range of the proposal extraction. Then, the EdgeBox algorithm is improved to make it suitable for the text proposal extraction in natural scenes. In addition, the proposed method is evaluated on two standard natural scene text detection benchmarks, and compared with other existing methods. Results show that the proposed method has better performance and robustness than other methods.

Abstract:With the development of spread spectrum technology, the binary offset carrier (BOC) modulation technology is utilized to improve the signal quality of the application system, reduce the signal mutual interference of the same frequency, and improve the security and anti-jamming. The key technology of side peak inhibition is studied, as the special and unique signal characteristics of correlation multi-peak and ambiguity discrimination caused by the new applied BOC modulation technology. Thus a moving correlation function acquisition algorithm is proposed, considering the limitation and deficiency of the existing acquisition methods. In the algorithm, the single peak correlation function is structured by moving correlation results to inhibit side peaks. Theoretical and simulation results demonstrate that the new algorithm can inhibit side peaks in the acquisition process, and it can adapt to lower SNR. In addition, this algorithm is better than the traditional algorithms in acquisition performance and inhibition side peak ability. The research provides the technical foundation to the efficient spread spectrum signal synchronization and system application.

Abstract:Firstly, from the perspective of subspace alignment, the joint optimization problem of the interference signal power and the useful signal power is modeled on the Grassmannian manifold. The constrained optimization problem is transformed into the unconstrained optimization problem with lower dimension. Then, using the geometric properties of the Grassmannian manifold, a joint interference alignment precoding scheme based on conjugate gradient algorithm on the Grassmannian manifold is proposed. Computer simulation results show that the proposed scheme improves the sum rate performance of the multiuser MIMO interference system by jointly considering the minimization of the interference signal power and the maximization of the useful signal power, and also improves the convergence speed by effectively solving the 90° turning problem of the Grassmannian steepest descent algorithm.

Abstract:For facial component warping and face makeup, the paper proposes a method of componentbased face editing and beautification. The method mainly includes two parts: Facial component warping and face makeup. It firstly carries out triangulation according to feature points of the components, then only changes the positions of feature points which need to be changed and finally calculates pixel values by interpolation. The warping results of different components are obtained. After the combination of specific warped images in different components, face beautification is realized using digital makeup based on example. Experimental results show that the beautified face looks real and more beautiful obviously. In short, the face beautification technology will bring considerable application prospects in the field of cosmetic surgery or cosmetic.

Abstract:Full-duplex model can transmit data during one time slot and frequency, which doubles the transmission data rate of relay systems compared with half-duplex. To improve the transmission rate of full-duplex multi-input and multi-output(MIMO) relay system with amplify-and-forward(AF) strategy, a beamforming scheme of maximizing rate is designed using alternating iterative structure and gradient descent method (GDM). We derive the beamforming expressions at relay station, and design an minimum mean square error(MMSE) iterative algorithm as the initial condition. Simulation results reveal that the proposed alternative iterative structure converge quickly and is better than some existing algorithm like zero-forcing(ZF), minimum mean square error(MMSE), maximizing signal-to-leakage-plus-noise ratio(Max-SLNR) algorithm.

Abstract:The emergence and development of high dimensional big data streams have presented a great challenge to the traditional machine learning and data mining algorithms. Based on the characteristics of data flow, first we construct an adaptive incremental feature extraction algorithm model. Then, according to the environment with noise, we establish an incremental manifold learning algorithm model based on feature space alignment to solve the small size sample problem. Finally, the regularization optimization framework of manifold learning is constructed to solve the problem of dimensionality reduction errors of high-dimensional data flow in feature extraction process, and then the optimal solutions are obtained. Experimental results show that the proposed algorithm framework conforms to the three evaluation criterions of manifold learning algorithm: Stability, enhancement, and the learning curve can rapidly increase to a relative stable level. Thus the efficient learning of high-dimensional data streams can be realized.

Abstract:k nearest neighbor (kNN), which is one of the most typical data mining algorithms, is widely applied in various areas due to its better generation ability and sufficient theory results. The method needs to compute the distances between the test instances and all the training instances during executing prediction. However, it costs substantial time as facing the large-scale data. To solve the problem, we propose an acceleration algorithm for k nearest neighbor classification based on stratified sampling (SS-kNN). In the method, SS-kNN firstly divides the instance space into several subranges with the same number of instances, and then samples instances from each subrange, finally judges which subrange the test instance sit and finds its nearest neighbors from this subrange. Compared with kNN and its variant based on the random sampling, SS-kNN could not only obtain the similar classification accuracy, but also accelerates the running time by an average of 399 and 16 times respectively.

Abstract:Coronal mass ejection (CME) is considered to be the main driving force of space weather disasters. CME often appears with many other solar activities. Since coronal dimming is associated most closely with CME, the detection of coronal dimming can help to forecast CME. Continuous development of CME observations brings increasingly large amount of data, whereas the efficiency of the detection algorithms of coronal dimming needs to be improved. Three parallel algorithms are presented for a semi-automatic dimming detection algorithm. They make the foundation for real-time detection of coronal dimming. We firstly introduce the existing work about the topic, then analyze a semi-automatic dimming detection algorithm, which to some degree, improves the efficiency of artificial recognition of darkened area, but does not meet the requirements of real-time detection. Based on the principle of parallelism of Matlab R2014a, three different parallel algorithms are presented from different aspects, i.e. data, distance-calculation and image-divide. Experimental results show that the parallelization based on image-divide is the most efficient one among the three algorithms.

Abstract:Considering that array thinning usually causes the increase of sidelobe level in beam pattern, an optimization algorithm based on improved wind driven optimization (WDO) algorithm and convex optimization is proposed to achieve sidelobe suppression in thinned arrays. To solve the problem of lacking a universal parameter selection scheme in original WDO algorithm, an improved WDO algorithm is given by combining WDO algorithm with Gaussian distribution. Then, the peak sidelobe level (PSLL) is taken into the semi-circular thinned array as objective function. To obtain a lower PSLL, the improved WDO algorithm is adopted as global optimization algorithm to optimize the distribution of elements in thinned array. At the same time, convex optimization is utilized as local optimization algorithm to effectively obtain the optimal coefficients of valid elements, which guarantees the perfect match of the elements′ positions and coefficients. Simulation results indicate that, under a given numbers of elements, the optimization algorithm could achieve sidelobe suppression effectively. Moreover, it has better global search capability and faster convergence speed. Therefore, the proposed optimization algorithm can be taken as an effective approach for thinned array design.

Abstract:The analysis of the influence of the array installation errors on the angle measurement accuracy of the phased array radar is important for the research and development of the high-performance phased array radar. The array installation errors-angle measurement accuracy model of the phased array radar is established. The angle measurement accuracy of the phased array radar with inclination error, role-angle error and azimuth normal error is studied by simulation respectively. The precision of the deduced concise formula arrivers accuracy level of 0.001° compared with the angle measurement accuracy error caused by array installation errors. The deduced formula and conclusion offers theoretical basis and engineering guidance for assigning the accuracy index of array installation and analyzing the accuracy out-of-tolerance problem of angle measurement.

Abstract:A nonlocal means method based on multichannel joint estimation for color image denosing is proposed, including two steps as color channel combination filtering and color channel fusion filtering. In the step of color channel combination filtering, the noisy color image is denoised by the classical nonlocal means of color(NLMC), from which the pre-denoised image is obtained as the input of color channel fusion filtering step. In the step of color channel fusion filtering, the pre-denoised image is denoised once more by generalized multichannel nonlocal means(NLM), and the similarity between the high frequency components of the pre-denoised image′s RGB channels is used in the denosing process at the same time. Experimental results demonstrate that the proposed method produces competitive results for both quantitative and visual comparisons with other classical color image denosing algorithms.

Abstract:This paper presents the study of posteriorgram features optimization based on nonnegative matrix factorization (NMF) algorithm and modified segmental dynamic time warping (SDTW) detection for unsupervised query-by-example spoken term detection. First, a Gaussian mixture model (GMM) is trained with frequency domain linear prediction (FDLP) acoustics feature parameters instead of Mel-frequency cepstral coefficients (MFCCs). Then the NMF algorithm is applied to the generated Gaussian posteriorgram matrix, and the derived base matrix is used as a subspace transform matrix for projection of raw feature. The projection can highlight the primary component of features and smooth the distance matrix. In the detecting phase, the best matching score is modified by using multi adjacent output scores, instead of the 1-best output score for normal SDTW. Experimental results show that without affecting detection time, the proposed method consistently outperforms the baseline systems with MFCCs and FDLP features with the detection precision improved by 18.6% and 18.1% respectively.

Abstract:Aiming at the problem of blind despreading of short-code BPSK/DSSS signals under non-cooperative and low signal-to-noise-ratio(SNR) circumstances,an algorithm which hierarchically deals with the signal is proposed.Firsly the sequence′s spreading period are estimated. Then the initial bit of a short-code sequence is figured out by employing the auto-correlation properties of short-code DSSS signals.Two windows whose length is spreading period are constructed at the same time,and by inner products of the two windows the relative polarities of all the adjacent information bits can be estimated.The robustness of the proposed algorithm is improved by correction of short-code sequence′s initial bits via a sliding method in every iteration. Simulation results show that a BER lower than 0.000 2 could be achieved under a SNR above -2 dB,which proves the engineering application value of the propsed algorithm.

Abstract:To improve the quality of K-Means clustering in highdimensional data, a K-Means clustering algorithm is presented based on non-negative matrix factorization with sparseness constraints. The algorithm finds the low dimensional data structure embedded in high-dimensional data by adding l1and l2norm sparseness constraints to the non-negative matrix factorization, and achieves low dimensional representation of high dimensional data. Then the K-Means algorithm, which is the high performance clustering algorithm in low dimensional data, is used to cluster the low dimensional representation of high dimensional data. The experimental results show that the proposed algorithm is feasible and effective in dealing with high-dimensional data.

Abstract:l1-minimization algorithm is one of the hot topics in the signal processing and optimization communities in solving the sparsest matrix. Compared with the traditional principal component analysis using l2 norm, the l1 norm only calculates the main characteristics matrix of the image, which is more robust to noise and abnormal data. While it is used too few in wood identification. The local binary pattern (LBP) texture analysis operator is defined as a gray-scale invariant texture measure. LBP algorithm is important in view point of pattern classification, and can be used to extract three-layer cross-sectional features of different wood RGB images data. And then a fast l1 norm algorithm is used to implement fast and accurate identification to judge whether the wood surface has defects or not, and where defects locate. Many experiments indicte that fast l1 algorithm combined with LBP can get correction of 0.931 for defect location in wood surface.

Abstract:There are three problems in rule set which is extracted based on classification algorithm.First, too few short rules in the extracted classification rule set decrease the number of high quality rules. Second, there are such few rules in rule set that almost all of the examples in the training data can be covered only once.Third, despite lots of extracted rules, some examples of small classes in the training data fail to be covered by any of these rules. Herein, a modified example multiple coverage classification algorithm RCIM, which is based on generated rules, is proposed. Here are the features: (1) for the purpose of improving the quality of rules, not only the quality of attribute value but also that of its complement can be taken into account when choosing the first item of a generated rule. (2) It can generate high quality rules at a time as many as possible. (3) It deletes the examples in the training data only if they are covered at least twice.What′s more, it can restudy each of the attribute value of the test data to extract rules when encountering the data difficult to judge.The algorithm RCIM not only can efficiently extract a large quantity of rules but also largely improve the quality of rules. Experimental results in many data show that RCIM has achieved higher classification accuracy than many other algorithms.

Abstract:To deal with the problem of inadequate sample in multiple bird species recognition, a new recognition method of multiple bird species in audio recordings is proposed based on feature transfer, which uses bird sounds of single species to train a multiple bird species recognition model. Maximum mean discrepancy (MMD) is used to measure the feature distributions difference of bird sounds, which maps audio feature of single-species bird sounds and multiple-species bird sounds into a new latent feature with the same distribution. Then single-species bird sounds with latent feature can be used to train a model of multiple-species bird sounds. The experimental result shows that method can achieve good regognition performance in a natural multiple-species bird sounds dataset based on four multi-label metrics. The recognition rate of proposed method increases by 20% compared with other methods in an artificial multiple-species bird sounds dataset.

Abstract:An improved compression tracking (ICT) algorithm is proposed based on the characteristics of aerial video. After study on classic compression tracking(CT) algorithm, some shortcomings in sample collection and classification of the sampling processing are found and improved. Kalman filter is used to predict the target movement path and the prediction results are applied to sample collection for adaptive research. The sampling and feedback of the classifier are updated by using the classification results after determined. Values lower than a certain threshold are not transfered to the classifier for classification, which ensures the correctness and the accuracy of the feedback of the classifier. Based on the proposed algorithm, a target tracking system is implemented for aerial video. Compared with the classic compression tracking algorithm in real aerial video, the effectiveness and real-time performance are tested and verified.