Multi-Polarization Sar Ship Fusrecognition Based On Deep Learning

Synthetic Aperture Radar(SAR)is able to image stably under all the time and all the weather,and thus plays an irreplaceable role in military and civilian applications.However,SAR images automatic target recognition is still a worldwide problem limited by the complexity of SAR scenes,target diversity,coherent speckle noise,SAR systems and observation platforms.In recent years,deep learning has made remarkable achievements in the field of optical image recognition.Related technologies have been introduced into SAR image recognition,and significant results have been achieved.However,most of the existing methods only use the information of single-polarized SAR images,and never fuse the complementary information of multiple polarization modes of SAR images.So there is still room for improvement in recognition accuracy.Around three key indicators of SAR automatic recognition system,recognition accuracy,inference time and training time,in this paper an efficient and accurate SAR recognition algorithm is designed,and the main work includes:1.A lightweight version Hourglass network is designed to predict the target’s heat map,and extracts the ROI of SAR images from the complex speckle noise background.The result of the experiment shows that the pre-screening network relieves the influence of Speckle noise of the background while ensures real-time performance of the SAR automatic recognition system.2.A dual-channel fusion recognition network is designed to integrate the complementary information between different polarization modes of SAR images,and greatly improves the recognition accuracy of the automatic recognition system compared to the way that only uses the single polarization modes.When Efficientnet-B2 is used as the backbone,our proposed architecture obtains an accuracy of 87.04% on the Open SARship dataset,which exceeds SOTA about 3%.In fact any classification network can be used as the backbone of the proposed architecture.3.A simple and efficient loss function,Fusion Loss,is designed to accelerate the convergence speed of the fusion recognition network.The experiments shows the convergence speed of DFRN accelerates75% averagely while ensuring the recognition rate.We prove the loss function works well on different optimizers by experiments.