| Synthetic Aperture Radar(SAR)is an imaging radar that is all-day,all-weather,and unaffected by clouds,fog,and weather factors.With the rapid development of SAR technology,research on object detection of ships in SAR images has become a very important research hotspot.The use of high-performance object detection algorithms to detect ships on the sea has important significance in various fields such as marine economic construction and naval deployment.In recent years,the detection performance of object detection algorithms based on deep learning has been continuously improved.However,in different ship detection scenarios of SAR images,due to issues such as unclear ship features,complex background interference,ship scale and dense layout,existing object detection calculation models have brought many difficulties and challenges.Based on the YOLOv5 algorithm,this thesis analyzes the problems encountered in SAR image ship detection tasks and improves the detection algorithm.The main work and achievements of this thesis are as follows:(1)This thesis proposes a target detection model based on multi-scale feature fusion.In response to the problem of large changes in the scale of ship targets and the majority of small targets in SAR images,this thesis proposes to add a small target scale detection layer to the YOLOv5 network,and designs a Bi F4 feature fusion network with four detection scales to fuse more target feature information and enhance the ability of small target detection.This thesis introduces the Swin Transformer module to improve the C3 module and obtain the C3 STR module,and experiments are conducted to determine the optimal replacement position of the C3 STR module,thereby improving the model’s perception of target feature information.Finally,in response to the situation where there are many targets densely arranged in the nearshore scene,this thesis uses the Gaussian weighted Soft-NMS to perform pre-selection box filtering work.The experimental results show that the detection accuracy of the model has been greatly improved by using three improvement points simultaneously,and the cases of missed and false detection of ship targets have been reduced.(2)This thesis constructs a SAR image ship dataset using rotation box annotation.This thesis enriches the experimental data sources of rotation detection models by slicing large scene SAR images obtained from SAR satellite image data,and then labeling ship targets in the slicing images with rotation boxes.By combining optical remote sensing to assist in labeling rotation boxes in complex scenes,the effectiveness of the rotation box model is fully verified through comparative experiments and generalization performance analysis experiments on this dataset.(3)This thesis proposes an object detection model that can use rotation box detection.In view of the characteristics of ship targets in SAR images such as large aspect ratio and strong directionality,this paper proposes to use CSL circular smooth tag technology to make the model learn the rotation angle through classification methods,and redesign the loss function,so that the model can realize the rotation detection function.Experiments have shown that using circular smooth labels on the YOLOv5 model or the improved YOLOv5 model to achieve rotation detection function has improved the detection accuracy of the model and the representation ability of ship targets. |
