Research On Target Recognition And Classification Of Sar Image Based On Deep Learning

Synthetic aperture radar(SAR)imaging systems have its unique advantages over optical imaging systems and can work at any time under different climatic conditions.In recent years,both in the military and civilian fields,the application of SAR images has increased continuously,and the requirements for SAR target recognition technology have become higher and higher.As the deep learning theory becoming research hotspot for artificial intelligence researchers,deep learning models have been continuously proposed in the field of image processing and other areas,and have achieved remarkable success.In this paper,based on the characteristics of SAR images,relying on the convolutional neural network(CNN)model in deep learning,the classification and recognition methods of SAR images based on deep learning is studied.The main research contents are as follows:The main characteristics of SAR images,including resolution and statistical distribution characteristics are described in this paper.The main difficulties in SAR image target recognition are analyzed.That is,when applying traditional identification methods,a large amount of prior knowledge is needed,and humans lack acknowledgement of target features in SAR images,so it's hard to select target features in SAR images effectively.On account of the characteristics of blind learning and unsupervised learning in deep learning,the CNN model is utilized to solve this problem.Because deep learning models are very sensitive to parameters,the effects of parameters such as activation function,pooling method,kernel size,batch size,and the number of convolutional layers on network performance are analyzed in this dissertation.Research results indicate:1.The performance of the network is affected by different activation functions.The commonly used ReLU activation function can be effectively replaced by swish activation function.2.For SAR targets,average pooling outperforms max pooling.3.Different kernel size and batch size have certain influence on SAR target recognition rate and network computing efficiency.4.A larger number of convolutional layers can improve the recognition rate for SAR targets.In this paper,the optimal values of various parameters are selected according to the experimental results.The recognition rate for SAR targets can reach over 99%.For SAR images in different backgrounds,the effects of different background factors on SAR target recognition accuracy are analyzed in this thesis,and two methods to reduce such effects are proposed.One is to extract the target and eliminate the influence of background factors as much as possible;the other is to expand the training data set and increase the number of SAR images samples in different backgrounds.Experiments show that both methods can effectively reduce the influence of different backgrounds on SAR target recognition.At the end of the thesis,the influence of the shadows on SAR target recognition is studied.In the absence of similar shadow training samples,there are a significant drop in recognition performance.Therefore,in order to improve the SAR image target recognition rate under complex conditions and improve the recognition performance of the system,we can increase the number of training samples under similar conditions.