Deep Curvelet Network For SAR Target Recognition

The development and popularization of Synthetic Aperture Radar(SAR)is derived from last Century,its main function is to track aircraft and warships and other military targets.SAR target recognition is the last part of SAR image processing and interpretation,playing a key role in military affairs.Deep learning algorithm has been popular in the field of image processing,and Master in the field of chess has achieved remarkable success,reflecting the deep learning algorithms has great advantages in the field of artificial intelligence.The curvelet is a kind of wavelet function with directional information.At the same time,it is also a multi-scale analysis tool,which can effectively capture the high-dimensional structure of data.The combination of deep learning algorithm and curvelet have great advantage in the field of SAR target recognition.This thesis mainly includes the following three aspects:Deep curvelet network is proposed for SAR target recognition in this paper.Firstly,the curvelet dictionary was constructed,the curvelet atom as the first Layer of the convolutional network,applying it to SAR target recognition.The deep curvelet network combines the multilayer structure of the convolutional network and Multiscale and multiresolution properties of curvelet,overcoming the influence of SAR image speckle and background clutter and the direction angle of the barrel.The precision of the recognition is developing,and the convergence of the network is accelerated.Multi-scale online Elastic-net and Multi view deep convolutional network.was presented for SAR target recogintion.This method mainly uses the ELastic-net to carry on the sparse coding to the image,then uses the max pooling to subsample the coefficient,inputing to the network.Because the single scale is not accurate to describe the detail of the image,the multi-scale method can get the structure information of the image.The multi-scales coefficients are input into multi-channels of the network to be identified.Finally,the target recognition of SAR image is highly sensitive to the attitude angle of the target,the method of multiview is introduced,which makes the data have nothing to do with the angle and only effecting it by the target scattering center,achieving a good SAR target recognition Hierarchical curvelet recurrent neural network was proposed.This method makes full use of the relationship of context of the image,overcoming the convolution network without considering the spatial dependence of the image.The curvelet was the first layer of the network,combining with a two layer LSTM network.The first layer of the LSTM is the coding layer and the second layer is the decoding layer.The proposed method can fully exploit the context relation and spatial position relation of SAR images,and finally has a good effect on SAR target recognition.