Abstract:Image super-resolution reconstruction is an image processing technology, which recovers high-resolution images from low-resolution images. While, the super-resolution problem is under-determined. In recent years, researchers have proposed learning-based methods to learn image prior information from a large amount of data, in order to constrain the super-resolution solution space. This paper introduces the mainstream image super-resolution reconstruction algorithms in the past two decades, which are divided into two categories: traditional features based methods and deep learning based methods. For the traditional super-resolution reconstruction algorithms, this paper mainly presents the methods based on neighborhood embedding, the methods based on sparse representation, and the methods based on local linear regression. For the deep learning based methods, the super-resolution model design, the up-sampling method and the loss function form are provided. In addition, this paper introduces the application of super-resolution reconstruction technology in video super-resolution, remote-sensing image super-resolution, and high-level vision tasks. Finally, the future development directions of image super-resolution reconstruction technology are provided.

Abstract:Facial expression recognition is the basis of human emotion recognition, which has been a hot topic in pattern recognition and artificial intelligence. This paper summarizes the development process of facial expression recognition framework, mainly including the traditional expression feature extraction, expression classification method and deep learning-based expression recognition method, and then analyzes and compares the recognition rate and performance of various algorithms. Moreover, this paper introduces the commonly used datasets of facial expression recognition and the advantages and problems of each data set. In view of these problems, the techniques and methods for data enhancement are analyzed, such as generative adversarial network (GAN). Finally, the existing problems in the field of facial expression recognition are summarized and the prospect of future development is put forward.

Abstract:In recent years, with the continuous development of computer vision technology and digital image processing technology, flame detection technology based on video analysis has attracted more and more attention. Aiming at the practical significance of video-based flame detection for fire safety, people and national property safety, and the diversity of theoretical basis and influencing factors, the existing video-based flame detection algorithms are summarized. Firstly, this paper introduces the current processing flow of video-based flame detection technology, that is, pre-processing, feature extraction, classification and recognition; secondly, the flame characteristics are analyzed including static features based on single frame and dynamic features based on multi frame, and the typical flame feature extraction algorithms are enumerated. Then, the multi feature fusion, and the classifier classification and the depth learning recognition method for the flame recognition are emphatically summarized. Finally, the difficulties and future development of video-based flame detection technology are discussed in detail.

Abstract:In order to achieve efficient, high precision and inexpensive full view 3D reconstruction, a method of full-view 3D reconstruction by fusing depth camera and illumination constraints is proposed. In the single frame reconstruction, a method of 3D reconstruction by fusing RGBD and shape from shading (SFS) is used, that is, the illumination constraints are added in the original depth data to optimize the depth value. In the registration of two adjacent frames, fast point feature histogram (FPFH) features are used for matching and filtering out the wrong matching points by random sample consensus (RANSAC) algorithm. Then the transformation relation between cameras is obtained through iterative closest point (ICP) algorithm. In the full angle of 3D reconstruction, the bundle adjustment is used to optimize the position and pose of the camera in order to solve the problem that the first and last frames can not be completely overlapped by the cumulative error. Finally, a complete model is generated. The method integrates the illumination information of the surface of the object, therefore, the generated 3D model is smoother, and contains more detailed information of the surface of the object, which improves the reconstruction accuracy. The method can complete the reconstruction of multi-reflectivity 3D objects in a natural light environment with a single photo, and has a wider application range. The entire experiment can be carried out with a handheld depth camera, which makes it easier to operate without turntable.

Abstract:In order to restore high-quality and sharp image from motion blurred images, the coded exposure imaging technology is adopted. Different from the shutter keeping constant open in traditional camera, the shutter of coded exposure camera is in the conversion between on and off. The high frequency information of the target is effectively preserved, since the fast conversion of code in time domain is equivalent to broadband filter in frequency domain. To obtain the sharp image from the image by coded exposure, image restoration and blurred kernel estimation are recovered based on L ₀ regularization, which can preserve the high frequency details. The sharp image is rebuilt by iterative updating of the restored image and the blurred kernel. Experiments on synthetic and actual images show that the proposed method has good image restoration effect for deblurring caused by various motions.

Abstract:In the vector of locally aggregated descriptors (VLAD) method for image representation by residual accumulation, the residual values obtained by each descriptor and the corresponding nearest neighbor codeword are different, and the number of descriptors corresponding to each codeword is uncertain. Thus, there are over cumulative and under cumulative problems. In this paper, a new method for image representation by residual center aggregation of distance clustering is proposed. First, the local descriptors of the database image are extracted, and the codebook is obtained by clustering these descriptors; then, the local descriptors are quantized to the codebook by the nearest neighbor method, and the Euclidean distances between the local descriptors and the corresponding nearest neighbor codeword are obtained. Again, after clustering all the distances and obtaining the central set, the method finds the nearest neighbor of the Euclidean distance between each local descriptor and the nearest neighbor codeword on the central set, obtains the descriptor corresponding to each center in the central set, determines the center of the residual between the descriptor corresponding to each center in the center set and the nearest neighbor codeword, and accumulates and summarizes all the residual centers on each codeword. Finally, all the cumulative vectors corresponding to the codewords are cascaded in order to get the final image representation. The results of image retrieval experiments on the Holidays and UKB datasets show that the proposed image representation method is better than the VLAD method by directly accumulating residuals and performing image representation.

Abstract:In the framework of the traditional isotropic total variation (ITV) denoising algorithm, the edge of the image is easily blurred, and it is difficult to maintain the image detail information. Thus, we mainly study the image denoising problem by using L _p pseudo-norm and ITV. The L _p pseudo-norm takes place the L ₁ norm in traditional ITV energy function. Then the energy function is broken into several sub-problems by using alternating direction method of multipliers (ADMM). We treat the differential operator as a convolution operator, then introduce the convolution theorem and the fast Fourier transform (FFT) to further improve the efficiency of the proposed method. Finally, through Matlab, the simulation experiment employs the objective and subjective evaluation methods of image quality for evaluation and analysis. Simulation results show that the proposed method is capable of preserving the edge characteristics of the image and improving the denoising effect efficiently.

Abstract:Wireless medium access control (MAC) protocol identification based on power detection plays an important role in cognitive radio and cognitive electronic warfare. To improve accuracy, features like collision proportion estimation and Fisher statistic are derived. Aiming at the problem of low accuracy when the target network is different from training samples, the selective ensemble method based on Q-learning is utilized. The identification system is constructed from a set of selected base classifiers to improve generation capacity. The proposed method is tested with samples of four MAC protocols by OPNET simulation. Experimental results show that the features proposed in this article increase discrimination ability of different MACs. Moreover, when the target network is different from training samples, the selective ensemble method has higher accuracy than the single classifier and the ensemble of all classifiers.

Abstract:In this paper, we propose a rank-reduced Capon algorithm for parameter estimation of near-field sources with uniform linear array. The proposed algorithm can simplify the two-dimensional (2D) peak search within the conventional 2D-Capon algorithm to one-dimensional peak search, which significantly reduces the computational complexity. In addition, the proposed algorithm can obtain automatically paired angle and range estimations of near-field sources, and the parameter estimation performance of the proposed algorithm is very close to the conventional 2D-Capon algorithm. The simulation experiments indicate the effectiveness and superiority of the proposed algorithm.

Abstract:Traditional unmanned aerial vehicle (UAV) air-to-ground channel models cannot support three-dimensional (3D) flight trajectories and 3D antenna arrays. In this paper, by introducing the spatial rotation matrix and trajectory parameters, a modified 3D geometry-based stochastic channel model for UAV-based air-ground links is proposed. Considering UAV air-to-ground communication scenarios, the proposed model adopts the time-varying Boolean variables to describe the dynamic birth and death processes of line-of-sight path, ground reflection path and other scattering paths. Meanwhile, the corresponding time-evolving algorithms of channel parameters, i.e., two-dimensional angle, delay and power, are given and analyzed. The upgraded method can be used to reproduce the time-variant statistical properties of UAV air-to-ground channels. The numerical simulation results have demonstrated that the output spatial-temporal correlation characteristics and Doppler power spectrum of our model are consistent well with the theoretical ones, and the auto-correlation characteristic agree well with the measured one. The proposed model is helpful for communication system optimization, algorithm verification and performance evaluation of UAV-based wireless communication equipment.

Abstract:Unmanned aerial vehicle (UAV) communication is a research hotspot in the field of wireless communication. In order to ensure the reliability of communication between ground mobile terminal and UAV, a cooperative relay transmission scheme based on space-time coding (STBC) is proposed in this paper. Then, by using cognitive radio technology to improve the spectral efficiency, we formulate an optimization problem to maximize the throughput of cognitive relay network under the constraints of the quality-of-service (QoS) requirement of the primary users and the transmission power limitation of cognitive users with amplify-and-forward (AF) and decode-and-forward (DF) protocols. The Lagrange multiplier function and Karush-Kuhn-Tucker (KKT) condition are combined to solve the optimal power allocation. Finally, the simulation results not only verify the effectiveness of the proposed scheme, but also show that the throughput of cognitive relay network with DF protocol is better than that of AF protocol.

Abstract:The minimum mean square error (MMSE) detection method in the massive multiple input multiple output (MIMO) system has a problem that the matrix inversion complexity is too high. In recent years, there have been many studies to reduce the complexity. How to improve the convergence speed and detection performance of the algorithm while reducing the complexity of the detection algorithm has always been the focus of attention. The symmetric accelerated over-relaxation (SAOR) iterative algorithm is applied to the signal detection of massive MIMO systems, which avoids complicated matrix inversion calculation, and the implementation complexity is reduced by an order of magnitude compared with the MMSE method. The simulation results show that the SAOR-based detection method can approach the detection performance of the MMSE algorithm with fewer iterations, which provides a better implementation method for the fast detection of received signals in massive MIMO systems.

Abstract:A kernel ridge regression (KRR) based localization algorithm is proposed for localization in wireless sensor networks. KRR algorithm adds kernel function on the basis of ridge regression. In the off-line phase of the algorithm, the KRR method is used to extract the nonlinear relationship between fingerprint data of all positions, and the nonlinear regression positioning model is trained. The received signal strength indicator(RSSI) value of the target is collected in the online stage, and the target position is estimated using the positioning model. Simulation analyzes various factors which of impact the algorithm performance. The location experiment is achieved in indoor typical office environment. Experimental results show that under the influence of different factors, this algorithm can achieve better positioning accuracy than the traditional WKNN algorithm. When the position grid spacing is 1.8 m, the average positioning error of WKNN algorithm is 2.53 m, while the error of the proposed algorithm is 1.58 m.

Abstract:At present, aiming at the problem of many trapping sets are generated when applying the traditional method to construct (3,m) LDPC codes,this paper proposes an improved scheme to construct (3,m) LDPC codes,which eliminate elementary trapping sets(ETS) based on rectangle lattice.By analyzing the relationship between ETSs and slopes,we can select the proper slopes which satisfy the corresponding constrained condition so as to avoid the emergence of ETSs;Simultaneously,quasi-cyclic structure is applied to the parity-check matrix in optimized scheme,which reduces the complexity of encoding and decoding for LDPC codes.The results of simulation experiments on the AWGN channel demonstrate that the proposed constructional algorithm in this paper can reduce the error floor effectively.

Abstract:Aiming at the problems of poor fault tolerance and low recognition rate of the existing LDPC code verification vector recognition algorithms, an iterative recognition algorithm for LDPC code verification vector is proposed. Firstly, the probability distribution of the log-likelihood ratio of the parity relation is analyzed by using the soft information of the channel output, and a pre-decision method of the parity vector is found. The theoretical values of the pre-decision threshold and the size of the data matrix are deduced, which greatly reduces the search space for constructing the parity vector. Then, under the condition of soft decision, using the generalized logarithmic likelihood ratio, the vector in the search space is judged and the check vector is obtained. Finally, multiple sets of data are used to iterate to find the check vectors in the search space. Experimental results show that compared with the existing algorithms, the vector search space of the proposed algorithm is greatly reduced, and the recognition rate is also significantly improved in the low SNR environment. Moreover, the decoding gain of the algorithm is increased by about 2.3 dB.

Abstract:To solve the problems of wire network wiring interference, high equipment cost and meet the needs of large data storage and diversified data access, a system based on the internet of things (IoT) wireless network and IoT cloud platform is proposed for the remote fault diagnosis of machine tools. The system model is divided into four levels: acquisition layer, transport layer, arithmetic level and application layer. The data acquisition method based on object linking and embedding for process control (OPC) and multi-sensor fusion is used to obtain the running data of machine tools in acquisition layer; the transmission layer based on narrow band internet of things (NB-IoT) wireless communication technology and IoT cloud platform make the remote transmission, communication and storage of data come true; and the operation layer establishs diagnostic algorithm mode which combined BP neural network with expert system. Taking a machine tool spindle servo as an example, its fault phenomena is analyzed and the fault samples are obtained. The model is analyzed by error simulation, and the predicted results coincide with the expectations, which verifies the validity of the model.

Abstract:Satellite navigation receiver has two different underlying architectures, including scalar tracking and vector tracking. The vector tracking receiver processes all the channels using a center navigation filter. This architecture could utilize the sharing information for improving the receiver performance. However, the channels will affect each other in this architecture. Channels with signal blockage or weaker signal will affect the navigation filter operation, and it is necessary for carrying out channel status monitoring. In this paper, a long short term memory-recurrent neural networks (LSTM-RNN) is proposed and applied in the channel status monitoring. The innovative sequence of the center navigation filter is employed as the input vector of the LSTM-RNN. Simulation results show that the proposed method could detect faults effectively and ensure the positioning accuracy of vector tracking receiver.

Abstract:Rotator cuff tear is a kind of frequently-occurring disesae,and ultrasonography is the prefferred examination method for the disease,but the organization of hematoma in the nidus leads to false-negative in most cases.A new nonlinear approach based on statistical analysis of ultrasound radio-frequency(RF) signals is developed for identifying the lesion to improve the diagnostic accuracy. Firstly, the raw RF data is scanned and saved in the second harmonic mode,then the region of interest(ROI) and reference region (RR) are selected in one frame by applying the band-pass filter. The root-mean-square value of second harmonic in one or more cycles within ROI and RR are calculated and compared by Kolmogorov-Smirnow test for the p-value.The relative p-value is estimated by comparing the ROI and RR. The proposed method solves the problems of high subjectivity in identifying complex tissue and requisiting rich clinical experience in the ultrasound examination with accurate measurement and easy implement.