Research On SAR Image Target Recognition Based On Faster R-CNN Network

Synthetic Aperture Radar(SAR)is an all-day,all-weather imaging radar,and its SAR images can be applied to various fields.Therefore,the target recognition method of SAR image is a hotspot in the field of radar research.In addition,due to the tremendous improvement of hardware computing power,deep learning algorithms have recently begun to show prominence in various fields.After introducing the classical SAR target detection algorithm,this thesis introduces Faster R-CNN in the field of deep learning into the problem of SAR image target detection and recognition,and carries out the following work around this problem:1.Aiming at the lack of SAR image data,this thesis verifies that DCGAN(Deep Convolution Generative Adversarial Networks)can generate effective sample data sets by comparing the Barkhault distance,cosine value and cosine angle of the gray histogram of the generated sample and the original sample.DCGAN and traditional data enhancement methods are used to expand the basic data set.2.Aiming at the problem that it is difficult to set Super-parameters in deep learning,three pooling methods,six convolution kernel sizes,four activation functions,seven loss function coefficients and three Anchor sizes are studied in this thesis.Five of them are selected to carry out the follow-up experiments.3.Aiming at the problem that the conventional Faster R-CNN algorithm is difficult to train and slow to decline,this thesis proposes using Resnet instead of VGG network to complete the assignment of basic convolution network,and uses a new loss function to solve the problem of sample imbalance in the training process.The gradient disappearance problem in the network is reduced,the samples tend to be balanced,and the algorithm is more robust.4.Aiming at the problem that the conventional Faster R-CNN algorithm is difficult to recognize small targets,this thesis proposes a method of adding FPN(Feature Pyramid Networks)layer on the basis of the conventional model structure,which abandons the traditional pattern of single feature map and synthesizes the advantages of feature map at all levels.The validity of this method is verified by experiments.Finally,it is proved that the improved Faster R-CNN algorithm proposed in this paper can effectively recognize SAR image targets.Compared with the traditional Faster R-CNN,the target recognition rate is increased to 97.63%.