Study On The Anchor-free Method Of SAR Image Ship Detection Based On Multi-scale Feature Fusion

Ship targets are sparsely distributed in SAR images,and the design of anchor frame has a great impact on the accuracy and generalization of existing SAR image target detection method based on anchor.Therefore,a new method for ship anchorless detection in SAR image based on multi-scale feature fusion is proposed.Firstly,considering the diversity of the ship in SAR figure scale,and most of the test network is sensitive to small scale target capture is not enough,in the feature extraction stage design context information fusion module,access to the characteristics of each layer of context information,and feature level fusion layer in a bottom-up way,combined with the target area of context information,the characteristics of ships to be detected can be extracted more finely.Secondly,aiming at target prediction in complex scene framing the problem of insufficient accuracy,branch interaction module is put forward,in the test phase use classification branches optimization regression testing box,to improve the target frame’s precision,at the same time the new IOU branches on the classification,using test box with true value border occurring simultaneously ratio,improve confidence level positioning calibrate test box classification,Suppresses low quality detection boxes.When verifying the detection performance of the algorithm in this paper,the standard precision index of average precision and speed index of FPS are selected.The experimental results show that: On the SSDD and SAR-Ship-Dataset data sets,the proposed method achieves good detection results,with average accuracy reaching 92.56%and 88.32%,respectively.Compared with other methods for Ship detection in SAR image,the proposed method not only has excellent performance in accuracy,but also has no complicated calculation related to anchor frame.The speed index FPS of the method in this paper reaches 28 frames per second,which is relatively fast in detection speed and has certain practical significance for real-time target detection in SAR images.There are 26 figures,7 tables and 63 references in this paper.