Research Of SAR Target Recognition Methods Based On Deep Learning

Because of its unique advantages, Synthetic Aperture Radar(SAR) has been becoming one significant way of acquiring information in modern society and playing a crucial role in both military and civilian areas. As one way of acquiring information from SAR, the target recognition of SAR images is one of the research hotspots all the time.During recently years, there exists a new wave of research into Artificial Intelligence(AI) since the deep learning has been proposed. With the combination of unsupervised and supervised learning, deep learning has made a large amount of data without labels valuable of learning and in result made unprecedented achievements in field of target recognition, but lots of problems still need to be solved.Firstly, this paper summarizes the technology of target recognition of SAR images based on machine learning and displays the recognition results of MSTAR dataset through two methods, neural networks of supervised learning and principal component analysis of unsupervised learning respectively.Secondly, this paper points out the limitations of method of machine learning in target recognition, which means that it requires much professional knowledge and can’t extract features automatically which can represent the SAR targets. Based on this point, this paper proposes that the models of deep learning could solve these problems. Two models, Deep Belief Networks(DBN) and Convolutional Neural Networks(CNN) respectively, are then applied in target recognition of SAR images. The influence of parameters into the performance of these two models is analyzed and the typical values of these parameters are then shown when recognizing SAR targets.Lastly, since the existence of great amount of speckle noise, which is one of the crucial factors that affect the recognition performance of model, this paper compares the performance of two methods of Lee filtering and wavelet transformation in suppressing the speckle and comes to conclusion that improvement in recognition performance could be achieved when Lee filtering and two level wavelet transformation are combined based on above result.