Research On Target Detection In Sar Images Based On Anchor-Free Mechanism

SAR,also known as synthetic aperture radar,has the characteristic of strong penetration compared to other radar systems.It can see through vegetation on the ground and clouds in the air,and has no requirements on the climate and light conditions of the observation site.With the rapid development of neural network technology,the application value of target detection algorithms based on deep learning in synthetic aperture radar image scenarios has gradually attracted the attention of researchers.The performance of common anchor box-based target detection algorithms is usually affected by the artificially set anchor box scale and other hyperparameters.Especially in SAR,where the number of targets is small and the image size is large,the performance loss caused by this effect is magnified.The content of this thesis is the SAR image target detection algorithm based on the anchor-free mechanism,and it makes contributions to solving the problems in SAR ship detection and proposing improvement measures.First of all,multi-scale targets are often encountered in SAR ship detection tasks,and the detection results may have the problem of missed detection of small and medium-sized targets.In response to this problem,this thesis proposes an anchor-free improved network MFE-FPN based on multi-scale feature balance.While introducing an attention mechanism,it fuses and balances the upper and lower layers of the feature pyramid to improve detection accuracy.Secondly,due to the difference in the tilt angle of the SAR ship target,the target label frame usually also has a difference in length and width.Common anchor-free detection algorithms do not pay attention to this when designing networks.In this regard,this thesis proposes an anchor-free improved method DW-ATSS based on a label assignment strategy.This method solves the problem that the horizontal and vertical offset weights between the sample to be allocated and the center point of the truth box are different.The effectiveness of the above improved algorithms has been verified by the SAR ship detection public dataset HRSID.Finally,in order to combine the proposed improved method with actual production,this thesis designs and implements a SAR image interpretation software based on C/S architecture,which includes many practical functions such as registration and login,image browsing,training and detection,and data management.The implementation of the software simplifies the workflow of SAR image interpreters and improves work efficiency.