Study On Ship Target Detection Algorithm In SAR Image Based On YOLOv4

In recent years,deep learning algorithms have been widely used in various fields of computer vision and have achieved significant results.Traditional Synthetic Aperture Radar(SAR)ship target detection for SAR images has also introduced deep learning algorithms.However,mainstream deep learning target detection algorithms are only suitable for ordinary life images,and these models are large in size,high in complexity,and slow in detection speed.To address these issues,this thesis studied the SAR ship target detection algorithm based on YOLOv4,and the main work is as follows:(1)Use the YOLOv4 model to perform ship target detection.In the experimental and theoretical verification process,it was found that the original YOLOv4 network has poor performance in ship target detection.Therefore,on the basis of the original network,the ASPP module was first added to amplify the deep level features,and then the RFB module was used to expand the receptive field of the main feature extraction network,allowing more location and category information to be obtained for the shallow level features,In addition,a bidirectional feature fusion pyramid structure with adaptive feature fusion is designed,which can better fuse different information between deep and shallow levels and reduce target conflicts.Finally,the YOLOv4-ship model is proposed for ship target detection.Experimental results show that YOLOv4-ship can greatly improve the accuracy of ship detection,and the average accuracy on HRSID and SSDD datasets reaches 89.52%and 90.94%,respectively.(2)In response to the problem that YOLOv4-ship and other mainstream models cannot meet the real-time detection of military and civilian ships,YOLOv4-ship is designed to be lightweight.First,the lightweight backbone network Mobile Nets series were extensively studied and their principles were understood.Secondly,the ECA attention module was selected to improve the average accuracy and recall rate of the network.Then,the convolution kernel of the RFB module was improved to make the parameters less and more suitable for the ship length-to-width ratio.Finally,the lightweight ship target detection model Faster-ship is proposed,which can achieve realtime requirements while maintaining high detection accuracy,and the detection speed is increased by 7 times and 5.5 times compared to Faster R-CNN and YOLOv4-ship,respectively.