| With the rapid development of Synthetic Aperture Radar(SAR)imaging technology in recent years,SAR has provided mature high-resolution SAR images with rich semantic information in the field of Marine ship target detection.Among them,target detection is the focus of SAR image interpretation,which provides strong technical support for the construction of Marine monitoring and coastal defense.Based on deep learning target detection technology,this paper focuses on the detection performance of the deep learning algorithm under different ship features in SAR images,aiming to seek adaptive algorithms and improve the detection performance,accuracy,and credibility.The specific research contents are as follows:Given the problems of the range and azimuth ambiguity interference of SAR images ship detection,a method that verifies the robustness of the deep learning algorithm is proposed,and an adaptive improvement is made to the Faster-R-CNN algorithm in the presence of the interference.In this method,the robustness of the algorithm is mainly verified.By testing SAR images under different interference conditions,a more effective algorithm is selected to improve the stability.For the Faster R-CNN algorithm,the FPN network and Res Ne Xt network are used for full training and feature mining,and further improvement is made in multi-level and expansibility.The simulation results show that the robustness of the improved algorithm under ambiguity interference is better than that of the Faster R-CNN algorithm based on the Res Net network.It is proved that the robustness of the algorithm is significantly improved by multi-feature fusion and network expansion structure.According to the problems of SAR images ship detection with radio frequency interference(RFI),a method that evaluates the performance of deep learning algorithms is proposed;the Retinanet algorithm is also improved in the presence of the RFI.In this method,the RFI characteristics are analyzed and simulated.The corresponding test data sets are then obtained by adding different levels of interference,and the algorithm is comprehensively evaluated by setting corresponding evaluation indexes.Free Anchor mechanism and FPN network are added for Retinanet algorithm to enhance the processing ability of the algorithm in the adaptive prediction box.Simulation results show that,the improved method has a better performance than the original Retinanet algorithm among varieties of SAR images of ship target detection methods under RFI.Indicating the reliability of the adaptivity in terms of the stability of the algorithm and reflects the effectivity of the residual network,which can tolerate a certain value of RFI. |
