Abstract:As one kind of non-acoustic sensors,bone-conducted microphone has excellent anti-noise characteristics because its speech transmission channel naturally shields the influence of ambient noise.Recently,bone-conducted microphone has gradually played an important role in speech communication systems in various strong noise environments.However,due to the low-pass characteristic of human body and the limitations of sensor technology,bone-conducted speech sounds muffled and unclear. It is of great significance to improve the bone-conducted speech for more efficient speech communication in strong noise environments and wider application prospects of bone-conducted products.The blind enhancement algorithm means only the bone-conducted speech information can be acquired during the speech enhancement stage.This kind of direct enhancement algorithm is more applicable than the fusion technique with air-conducted speech.Here,the characteristics of bone conducted speech are firstly analyzed,then the existing methods including unsupervised bandwidth extension,equalization and spectral envelope transformation are introduced.Finally,we share some views of future research prospects.

Abstract:Extracting the characteristics of recording devices from audio signals is a frontier in judicial comparative research and audio forensics.As a research hotspot of audio forensics, recorder recognition technology is disturbed by environment, semantics, speaker and other factors. This paper introduces the development, basic theory and structure of recording equipment research. Especially, the research status of non-speech segment detection, feature parameters, recognition model and database construction is introduced and analyzed. Finally, the shortcomings and prospects of recorder identification are discussed. It is considered that the next stage should focus on how to speed up the database construction and the application of deep learning in recorder recognition.

Abstract:A speech enhancement algorithm based on deep belief network is proposed and improved for its shortcomings.Since there are few types of noise in the training set and the noise characteristics are not rich enough, the noise spectrum is disturbed in the frequency domain to enrich the noise spectrum characteristics. Considering that the signals of different frequency points have different effects on the system error, the weight coefficient is combined with the absolute hearing threshold. Finally, the better LOG minumum mean square error (LOG-MMSE) in the traditional speech enhancement algorithm and the improved deep confidence network-based speech enhancement algorithm in the noise environment are compared and analyzed. The result shows that the speech enhancement algorithm of the deep belief network exhibits excellent performance, especially the enhanced voice quality compared with the LOG-MMSE.

Abstract:Deep learning has become new state-of-the-art framework in many task in big data circumstance.Most of methods need full annotated data or assume only an object in the image with simple background.However,complex background,more than one object in the image and expensive full annotation in the reality,object recognition becomes more challenging.Here,we propose a deep-model-based attention mechanism and recurrent neural network.It trains the network end-to-end on multi-label data with image-level label.The glimpses change along with stochastic gradient descent and focus on different local region in every step.Finally,the effectiveness of the proposed algorithm is verified on the PASCAL VOC 2007 and 2012 datasets.Results show that the network is easily interpretable than other methods.

Abstract:It often plays a key role to extract homogeneous regions in the existing noise estimating methods for hyperspectral images (HSI). An effective homogeneous region detection method can improve the accuracy of image noise estimation. An isotropic homogeneous region detection algorithm (IHRDA) is proposed by using spatial information and spectral information, where a new Lance-SAD metric (LSM) is constructed to distinguish the similarity of picture elements in the homogeneous regions; then the noise level of hyperspectral images is estimated using the optimal regions with decorrelation based on multivariable linear regression (MLR) model. In experiments, synthetic images with different structure under different signal to noise ratio (SNR) and true hyperspectral remote sensing images are both compared with many existing methods, which show that the proposed method is more accurate and stable for hyperspectral images with various complexities and different noise levels.

Abstract:The performance of modern wide band software radio receiver (SDR) is deteriorated by nonlinear receiver with memory effect. Thus this paper puts forward a blind nonlinear equalization method called adaptive correlation elimination algorithm to suppress the nonlinear effect of Wiener receiver. Owing to independence of signals, taking the orthogonal frequency division multiplexing (OFDM) signal as an example, the algorithm detects correlation between signal and its remaining harmonic so that it can adjust the weights of nonlinear compensation model adaptively. The simulation result suggests that the algorithm can suppress the nonlinear effect of receiver under the large interference, reduce the error rate of weak signal, and enhance sensitivity of receiver, with low complexity.

Abstract:The problem of joint direction of arrival (DOA) and Doppler frequency estimation of coherent targets in a monostatic multiple-input multiple-output (MIMO) radar is addressed. Based on the propagator method (PM), an RD-FSS-PM algorithm is proposed, which can effectively estimate the DOA and Doppler frequency of coherent targets with low computational load. In the RD-FSS-PM algorithm, we firstly perform a reduced-dimension (RD) transformation on received signals to decrease the computational load, then use forward spatial smoothing (FSS) to decorrelate the coherent signals and apply the PM to estimate the DOA and Doppler frequency simultaneously, which are automatically paired. Compared with the conventional FSS-PM method, the RD-FSS-PM algorithm has much better DOA estimation performance, very close frequency estimation accuracy and much less complexity. The variance of the estimation error and the Cramer-Rao bound (CRB) of the DOA and frequency estimation are derived. Simulation results are presented to show the effectiveness and improvement of the new approach.

Abstract:The adaptation of transmitted waveform can enhance radar performances in detection, track, anti-jamming, etc. But the performance of this technique is linked with the prior knowledge of the radar target and clutters. In theory, the adaptive method of transmitted waveform based on maximum output signal-to-clutter-noise ratio (SCNR) criterion has been proved to be feasible and effective. In this paper, for enhancing practical application value of this adaptive method, the performance of the transmitted waveform adaptation method was verified by utilizing the prior knowledge of targets and clutters provided by Swerling statistic model and the digital elevation model (DEM) in airborne phased array radar. The simulation results show that the transmitted waveform can adaptively vary with the change of target and clutter for achieving the optimal matching of the two. Therefore, this method is better than the traditional methods in increasing output SCNRs of radar. It has good practical application prospects.

Abstract:How to improve the accuracy of face attributes recognition in natural environment or unrestricted environment is an important question in applying face attributes. In daily life, the uncontrollable factors, such as face postures and light, have a great influence on the recognition of human face attributes. How to improve the accuracy under the influence of the above factors is a key problem in the study of face attribute recognition. Given the success of convolutional neural network (CNN) in image classification, a new network structure is built by using multi-level sub-network and ranked Dropout mechanism algorithm. The structure has strong robustness to deal with face changes, thus achieving better results in the CelebA dataset and LFWA dataset, and reducing the network size significantly as well.

Abstract:Person re-identification is one of the important issues addressed in computer vision. Existing recognition system concerns the matching of pedestrians across over-lapping cameras. When assuming pedestrian images as one representation of the camera view, person re-identification can be considered as a multi-view learning problem directly. On the basis of this assumption, a pedestrian recognition algorithm is proposed via canonical correlation analysis. Since the canonical correlation analysis is a linear dimensionality reduction algorithm, it is hard to extract semantic information for person re-identification (such as low resolution of images, changing illumination and other factors). A sparsity learning based person re-identification algorithm (SLR) is proposed. First, SLR obtained the semantic information of each camera view by the sparse learning, and then mapped the high-level features t into a public hidden space in order to make characteristic distance between different views can be compared. SLR aims to obtain more discriminable public hidden space. Finally, improve the matching rate of person re-identification across disjoint camera views. Comparing the proposed method and other common methods on the VIPeR dataset and CUHK campus dataset, experimental results show that the proposed method has higher recognition efficiency.

Abstract:This paper proposes a fast algorithm for sparse decomposition of linear FM (LFM) signal to solve the deficiency of traditional methods that there are a large number of atoms. The atoms in the over-complete dictionary are structured based on linear FM signal, and fast sparse decomposition of LFM is completed by using combined dictionary. By analysis, the number of atoms in two over-complete dictionaries is much smaller than that in one over-complete dictionaries, and the most matched atom in another dictionary can be found with the use of searching the maximum based on fast Fourier transform. Simulation results show that the computational efficiency of the proposed method is better than that of three other algorithms using one over-completed dictionary, and the sparsity is better.

Abstract:For enhancing the target recognition effect of synthetic aperture radar image, a method of synthetic aperture radar image target recognition based on multi-linear principal component analysis and tensor analysis is proposed in this paper. Firstly, a four-order tensor training sample is constructed. Then, multi-linear principal component analysis is used to get the multi-linear projection matrix, and the core tensor is obtained from the multi-linear projection matrix. Finally, linear discriminant analysis is used to train the core tensor and classify the test samples. In the experiments, the proposed multilinear principal component analysis and tensor analysis method in this paper is applied to MSTAR public database for recognition experiments, and compared with principal component analysis and two-dimensional principal component analysis in recognition rate. Experimental results show that the method effectively preserves the image structure information and improves the target recognition rate.

Abstract:A method to eliminate baseline drift of ECG signal based on improved ensemble empirical mode decomposition is proposed for the disadvantage of poor filtering in traditional method. The method can weaken the mode mixing of empirical mode decomposition, and make up for the shortcomings of EEMD. It establishes the criterion for adding auxiliary white noise in EEMD method, and then determines the two important parameters,i.e., the magnitude of auxiliary white noise and the ensemble times. The method extracts the baseline drift signal from the noisy signal and then reconstructs intrinsic mode function to obtain the "clean" ECG signal, which provides a prerequisite for subsequent research. The experimental results show that the de-noising method, compared with the traditional method, can improve the SNR, reduce root mean square error, keep the characteristic of the waveform, and solve the problem of low frequency component loss.

Abstract:Technique for order performance by similarity to ideal solution (TOPSIS) is one of the multi-attribute decision-making methods and is used to solve supplier selections and other questions generally. In the process of evaluation focusing on production, there are many evaluation objects and hierarchical structures among objects. Decision makers pay more attention to the classification of objects, namely:good, medium or poor, though the evaluation systems for above problem are very insufficient. Therefore, this paper proposes a new three-way decision evaluation system based on TOPSIS. First, the analytic hierarchy process(AHP) method is used to calculate the weight of attributes. Then, to solve the hierarchy relations among objects, a preliminary classification is added before the process of evaluation. Finally, the improved TOPSIS method is used to evaluate objects. Case analysis result shows that the proposed method is effective.

Abstract:Link prediction is an important research direction of complex networks,and the method based on the node similarity is one of the most popular methods. So far, most of the node similarity prediction methods using link density have not considered the difference of each common neighbor node, that is, the contribution of different nodes to the link is different. Therefore, this paper proposes a link prediction algorithm based on the node contribution and link density of the common neighbor nodes(LDNC). The algorithm first calculates the link information between the common neighbor nodes as the link density of the nodes, and then defines node-coupling clustering coefficient to describe the contribution of the common neighbor nodes, and finally combines the two parameters.Experiments based on the real-world datasets show that the LDNC is more accurate compared with four baseline link prediction algorithms (CN,AA,RA and Jaccard) and the CNBIDE algorithm based on the node link density.

Abstract:To improve the discriminative ability of the coding coefficients, the Profiles (the line vectors of coding coefficients matrix) of Fisher discriminative dictionary learning (PFDDL) is proposed. Firstly, the Profiles can indicate the corresponding atoms which are used by the training samples to encode in the dictionary learning, and an adaptive method is proposed to construct the labels of atoms. Since there are one-to-one correspondences between the Profiles and atoms, then the Fisher discriminative criterion is imposed on the Profiles so that they have small within-class compactness but large between-class separability. Thus, it can encourage the atoms of the same class to reconstruct the training sample of the same class, and enhance the discriminative ability of the coding coefficients, then improve the performance of dictionary learning. Experimental results show that the PFDDL algorithm can achieve better classification performance than other sparse coding and dictionary learning algorithms on the three face and one handwriting databases.

Abstract:With the increasing popularity of social networks, sentiment analysis based on Twitter text has become a hotspot in recent years. The sentiment tendencies contained in tweets are important for mining user needs and predicting major events. However, the existing sentiment classification methods are mostly based on hand-made text features, and it is hard to mine implicit deep semantics of texts. In addition, because of special characteristics, such as short text and arbitrariness of users' behavior, it is more difficult to improve performance of current sentiment classification. This paper presents a novel Twitter sentiment classification model based on convolutional neural network (CNN). In order to explore sentiment tendency of tweets, the proposed model utilizes a dynamic CNN architecture to learn deep semantics from tweets, which initializes input word embedding with word2vec method. Experimental results show that our proposed model can achieve a recall rate of 82.3%, which is much higher than performances of traditional classification methods.

Abstract:A reconstructed 3D point cloud model based on image usually includes many outliers. These outliers may exist in isolation or dense aggregation forming point clusters, and they may also be distributed around the model or even on the mode surface.It is difficult to filter out outliers of various distribution states through one detection method. Therefore, a comprehensive outlier detection algorithm is proposed. Firstly, the outliers far from the main body of the model are eliminated by space distance, and the search speed of outliers is accelerated by constructing the spatial topological relationship.Then, the boundary matching method is used to filter out the outliers around the model by comparing the smaller clusters with the largest ones respectively. Finally, an improved K-means algorithm is used to cluster and classify the point cloud data according to RGB color value features, and the outliers adhering to the model surface are detected and filtered by combining the identified outliers.Simulation results show that the proposed algorithm can effectively filter out the outliers with multiple distribution states in the reconstructed point cloud.

Abstract:Class overlap is defined as the overlay degree of data from different classes, quantified by the approaches of geometrical statistics and information theory, and it is used to measure the complexity of a classification. There are imbalanced data in the real world, and the great disparity of the sample amounts challenges classification. With the help of experiments, we evaluate the efficiency of the class overlap measures on imbalanced data classification. Firstly, focusing on two-class classification, the experiments are designed to evaluate the efficiency of the class overlap measures on synthetic unbalanced data, which are generated with various skewness, class boundary shapes, feature types and probability distributions. Secondly, according to the experimental results on the artificial data, the influence rules of the imbalanced ratio on the measures are analyzed, then the ways of the measures to guide unbalanced data classification are concluded. Finally, the conclusions are evaluated on the real-world imbalanced data sets. The experimental results demonstrate that those measures with higher robustness on data skeness can efficiently guide classifiers selection for imbalanced data classification.