Deep learning is bringing revolution to pattern recognition and machine learning,which has been successfully applied to language processing,image processing,signal processing,business economy and so on.Restricted Boltzmann machine (RBM) is a strong representation and generative mod el,however,the learning time of deep belief nets (DBN),which consists of multiple stacking RBM,will be longer.In this paper,the improved momentum method is used not only in gradient ascent algorithm but also in gradient descent algorithm for both classification accuracy enhancement and training time decreasing.According to the characteristics of the gradient ascent algorithm,a rapidly ascending momentum method is used in the RBM pre-training phase,which greatly improves the speed of learning.According to the characteristics of the gradient descent algorithm,an improved slowly descending momentum term is also used in the fine-tuning stage to accurately find the best point.Through the recognition experiments on the MNIST dataset and CMU-PIE face dataset,the achieved results show that the improved momentum algorithm can effectively enhance the ability of image feature expression and improve both accuracy and computation efficiency.

To solve the problem of slow speed and low accuracy in the surface defect detection of hot rolled strips, an improved YOLOv3 algorithm is proposed. Firstly, the weighting K-means clustering algorithm is put forward to optimize priors anchor's parameters, which can improve the match between priors anchor and feature map. Secondly, the improved network structure of the YOLOv3 algorithm is proposed to improve the detection accuracy, whose shallow features and deep features are combined to form the new large-scale inspection layer. The experiments are carried out on the NEU-DET dataset, the results show that the average accuracy of the improved YOLOv3 algorithm is 80%, which is 11% higher than that of the original algorithm; the detection speed is 50fps, which is faster than other strip surface defect detection algorithms based on deep learning.

The bilinear convolutional neural network(B-CNN) has been widely used in computer vision. B-CNN can capture the linear correlation between different channels by performing the outer product operation on the features of the convolutional layer output, thus enhancing the representative ability of the convolutional network. Since the non-linear relationship between the channels in the feature map is not taken account of, this method cannot make full use of the richer information contained between the channels. In order to solve this problem, this paper proposes a kernelized bilinear convolutional neural network employing the kernel function to effectively capture the non-linear relationship between the channels in the feature map, and further enhancing the representative ability of the convolutional network. In this paper, the method is evaluated on three common fine-grained benchmarks CUB-200-2011, FGVC-Aircraft and Cars. Experiments show that our method is superior to its counterparts on all three benchmarks.

With the continuous expansion and complexity of network structure,the traditional overlapping community detection algorithm can not effectively discover reasonable community structure in large-scale network structure.Based on the concept of vertex gravity proposed in this paper,we introduce vertex cohesion and community cohesion as indexes for community structure-close internal structure and sparse external structure,and then put forward overlapping community structure algorithm OCSC.The steps of OCSC algorithm include pre-processing,core sub-mapping and core community expansion.Finally,NMI and F1Score confirm the rationality and superiority of OCSC algorithm by experimentation on synthetic and real network structures.

Aiming at that RGB image is rich in color details of scene and infrared image is sensitive to outline、size and boundary of target,a novel semantic segmentation model APFCN (Asymmetric Parallelism Fully Convolutional Networks) is proposed.In the upper part of APFCN,a five layer dilation convolution network,where the five kernel sizes are not uniform,is designed used to extract the high-level targets contour features of infrared image.In the lower part of APFCN,a classical CNN network is used to extract three scale features of RGB images.At the back of APFCN,the high level features of the infrared image are fused with the three scale features of the RGB image,and the fused features after 4 times upper sampling is used as the semantic segmentation output of APFCN.The results show that APFCN is better than FCN (input RGB image or infrared image) in PA (Pixel Accuracy) and MIoU (Mean Intersection over Union).APFCN is suitable for the semantic segmentation task of ground targets with consistent background colors.

Object detection is a hot topic in the field of computer vision, and has been widely used in robot navigation, intelligent video surveillance, aerospace, and other fields. The research background, significance and challenges of object detection were introduced. Then the object detection algorithms based on deep learning were reviewed according to two categories: candidate region-based and regression-based. For the candidate region-based algorithms, we first introduced the R-CNN (Region with Convolutional Neural Network) based series of algorithms, and then the R-CNN based methods were overviewed from four dimensions: the research of feature extraction networks, the region of interesting pooling researches, improved works based on region proposal networks, and some improved approaches of non maximum suppression algorithms. Next, the regression-based algorithms were surveyed in terms of YOLO (You Only Look Once) series and SSD (Single Shot multibox Detector) series. Finally, according to the current trend of object detection algorithms that are developing more efficient and reasonable detection frameworks, the future research focuses of unsupervised and unknown category object detection directions were prospected.

A cognitive radar target tracking algorithm is proposed for the tracking problem in complex battlefield environment.Based on the theory of human "perception-action" cycle,first,the Cramer-Rao lower bound (CRLB) of target radial distance,radial velocity and azimuth is approximated to the measurement error covariance.Then,the information entropy is used to describe the uncertainty of target tracking,and the connection between data processing in radar receiver and signal processing in radar transmitter is established with the criterion of minimum entropy.Furthermore,inspired by the three stage memory mechanism of human brain,"memory" is nested in Interacting Multiple Model (IMM) algorithm to overcome the tracking precision degradation problem when the model transition probability is set improperly.Thus,the transition probability can be adaptively adjusted to enhance the dominant model and weaken the bad competition of the mismatched model.The simulation results verify the effectiveness of the proposed algorithm.

Aiming at the problem that whale optimization algorithm is easy to fall into local extreme value and slow convergence speed,this paper proposes a whale optimization algorithm based on adaptive weight and simulated annealing.The improved convergence weight strategy is used to adjust the convergence speed of the algorithm,and the global optimization ability of the whale optimization algorithm is enhanced by simulated annealing.In the simulation experiment,18 test functions were calculated and the genetic algorithm,the particle swarm optimization algorithm and the standard whale algorithm were compared and statistically analyzed.At the same time,the influence of the adaptive weight and simulated annealing on the whale optimization is compared.The results show that the improved algorithm has a significant improvement in the calculation of the extremum of the test function,and the effectiveness of the improved algorithm is verified.

Collaborative filtering, as the core technology of recommendation systems, is currently facing the sparsity problem of rating data. This can be effectively solved through integrating item text information. However, current methods focus on extracting the one-dimensional features of the text, neglecting its multidimensional semantic features. Digging deeply into the multidimensional semantic features of the text can improve the recommendations. To help achieve this goal,a probabilistic matrix factorization model based on multidimensional semantic representation learning is proposed in the present study. The model uses a capsule network to mine the multidimensional semantic features of the text, and then integrates it into the probabilistic matrix decomposition framework using the regularization method to reveal hidden features linking users and items. Experimental results show that the proposed model has higher prediction accuracy.

Image super-resolution reconstruction (SR) aims to obtain high-resolution images from one or more low-resolution images. Recently, SR has been developing and widely applied in different fields. This survey retrospects the history of SR technique and provides a comprehensive overview of representative SR methods, with an emphasis on recent deep learning-based approaches. We elaborate the details of various deep learning-based SR methods, including their strengths and weakness, in terms of the deep learning model, architecture, and message pass. Finally, we discuss the possible research directions on SR technique.