SAR Image Classification Based On Convolutional Neural Network

Synthetic Aperture Radar(SAR),as a kind of microwave imaging radar,is widely used in economic and military fields because its advantages are all-weather and allweather.When it comes to related fields such as battlefield reconnaissance and precision strikes,SAR image target recognition technology is also used by people accordingly,and the need for SAR image target classification in various actual scenarios has also emerged.Usually target classification algorithms start from the characteristics of the data set,analyze the characteristics of different targets,and manually establish the corresponding feature extraction framework,which is heavily dependent on professional knowledge and is difficult to implement in practical applications.Therefore,deep learning methods began to combine with SAR images for classification.SAR images are often closely related to military secrets and other fields,and the cost of SAR imaging is high.These various reasons cause the existing only labeled SAR image data set to be relatively small,so the convolutional neural network cannot be directly applied to SAR.Image field.The main contents of this article are as follows:(1)A SAR image classification method based on convolutional neural network model migration is proposed.In view of the small sample size of SAR image data,first use a large number of ordinary optical image data sets to train on the deep neural network to obtain the pre-training model;secondly,transfer the pre-training model to the SAR image data set,and finally use a small sample of SAR The image performs local reinforcement training on important network parameters.(2)Propose a method to increase the sample volume of SAR image.First,the data set is processed by logarithmic transformation to obtain a data sample after data enhancement;in the SAR image imaging process,coherent speckle noise is always carried,and then the data set is processed by the filtering method with the best denoising effect,get another data sample after denoising.Finally,a multi-mode SAR image sample capacity increase method is established.(3)Propose a feature fusion SAR image classification method.According to the comparison of filter denoising effects,select a filter with good denoising effect to denoise the SAR image,and then obtain the denoised features;then use the edge detection operator to extract the edge feature information of the SAR image to obtain the edge feature;This feature is fused with the denoising feature,and the new feature map is used as the input of the convolutional neural network.