SAR Target Recognition Based On Convolution Neural Network And Migration Learning

Synthetic aperture radar has unique characteristics such as all-weather,wide band,all-day and active imaging,and has a wide development prospect in the fields of land and military security.However,it is difficult to identify SAR image targets manually with naked eyes,so it is necessary and necessary to automatically identify SAR image targets by using existing technologies.The main emphasis and difficulty of SAR image targets are SAR image preprocessing and SAR image feature extraction,This paper mainly discusses and studies SAR target automatic recognition combined with migration learning and convolution neural network algorithm.Firstly,this paper analyzes the basic concept and structure of SAR image automatic recognition algorithm,and introduces the characteristics of SAR image.Aiming at the unavoidable speckle noise denoising in SAR image imaging process,the commonly used denoising algorithms are analyzed,compared and tested.This paper discusses the data set preprocessing,feature extraction and classification and recognition of SAR image automatic motion recognition,which is a theoretical basis for subsequent work.Secondly,the algorithm of SAR image recognition by convolution neural network is studied.The convolution neural network VGG16 network structure is used to recognize SAR image targets enhanced by data,and the advantages and disadvantages of wavelet transform and Lee filter are compared experimentally.Finally,the experimental results show that the recognition rate and loss rate are obviously improved after data set denoising preprocessing,and Lee filtering preprocessing method retains more contour information in recognition rate and loss rate,which is better than wavelet preprocessing method.Finally,in SAR target feature extraction,migration learning is used to transfer the pre-training network model.By comparing the migration VGG16 with the convolution neural network experiment,the migration network structure can improve the recognition efficiency of the convolution neural network under the condition of ensuring the recognition accuracy.In order to determine the selection scheme of pre-training network structure,the migration VGG16 network and the migration ResNet network are simulated and demonstrated.ResNet network can solve the problem of gradient disappearance better than VGG16 network because of its deeper network layer number,so this algorithm adopts migrated ResNet network for SAR target recognition.The experimental results show that migrating ResNet network achieves higher recognition accuracy than migrating VGG16.