Deep Convolution Neural Network And Its Application In Ground Image Target Recognition

Deep learning is a new research field of artificial intelligence.Compared with the shallow machine learning method,deep learning simulates the hierarchical system of human visual nervous system which contains more hidden unit layers.Through layer-by-layer non-linear transformation of input data,a higher and more abstract features can be get.High-level features can enhance the discriminatory ability of the input data,while weak the adverse effects of unrelated factors.Deep convolution neural network is a main method of deep learning.By means of local connection,weight sharing and pooling operation,the deep convolution neural network parameters are reduced,the network complexity is decreased,and the invariance of target translation,distortion,rotation and scaling is reserved.At present,deep convolution neural network has made breakthroughs in many fields,especially in computer vision.However,there are still some core problems in the study of deep learning,such as gradient disappearance and model convergence.At the same time,the visual tasks will be more difficult because of the different illumination conditions,shooting distance and complex background environment of image acquisition.In order to improve the research foundation of deep convolution neural network and extend the application of deep convolution neural network,this dissertation studies the problem of gradient disappearance and model convergence in deep learning,and studies the tasks of image classification,object detection and semantics segmentation of ground targets under complex conditions.The main work and innovations of this dissertation are as follows:(1)Aiming at the problem of gradient disappearance of deep convolution neural network,the problem of gradient disappearance is analyzed by simplifying the structure of ResNet,and the back propagation process of the simplified model is deduced.At the same time,according to the changing characteristics of parameters in different layers of ResNet,an improved model A-ResNet based on adjustable shortcut connection is proposed.Experiments show that A-ResNet can reduce the gradient disappearance in deep convolution neural network.(2)Aiming at the high computational complexity of full parameter learning rate in deep learning,a combined learning rate strategy AdaMix is proposed for Autoencoder by analyzing the different effects of weights and biases.A full-parameter learning rate based on first-order gradient calculation is designed for weights,and a power-exponential learning rate is selected for biases.Experiments show that AdaMix can improve the convergence speed of deep learning model.(3)Aiming at the problem of low classification accuracy of ground image targets under the influence of illumination,a new illumination invariant extraction method MLNCST is proposed by analyzing the characteristics of high and low frequency components of LNSCT.At the same time,a parallel synchronous convolution neural network,Dual Lenet,is designed by fusing the high-level features of raw image and illumination invariant features.Experiments show that MLNCST and Dual Lenet can improve the accuracy of ground image targets under the influence of illumination.(4)Aiming at the problem of low detection precision in ground image targets with low pixel ratio,a new depth separable dilated convolution is proposed by analyzing the structural characteristics of SSD additional feature extraction network,and a parallel multi-scale additional feature extraction network including three independent self-networks is designed.Experiments show that the proposed network can improve the detection precision and speed of SSD for detecting targets in ground images with low pixel ratio.(5)Aiming at the problem of low semantic segmentation precision in ground image targets with uniform background,an asymmetric parallel semantic segmentation model APFCN is proposed by studying the characteristics of infrared and RGB images.At the same time,a dilated convolution rate strategy with non-fixed convolution kernel size is proposed to extract high-quality infrared image target contour information.Experiments show that APFCN can improve the semantic segmentation precision in ground image targets with uniform background.The purpose of this dissertation is to improve the performance of deep convolution neural network and study the recognition algorithms of ground image targets in complex environment based on deep convolution neural network.Those algorithms can be applied to UAV image recognition task,because it avoids the danger of close-range image acquisition,improves the ability of UAV to cope with extreme weather and complex environment,and increases the accuracy and speed of UAV image recognition.