基于跨域多焦点注意力改进YOLOv5s的SAR船舶目标检测算法  

Improved YOLOv5s SAR Ship Object Detection Algorithm Based on Cross Domain Multi Focus Attention

作　　者：胡宗仁 朱家兵 HU Zong-ren;ZHU Jia-bing(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan,Anhui 232001,China;School of Electronic Engineering,Huainan Normal University,Huainan,Anhui 232038,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001 [2]淮南师范学院电子工程学院,安徽淮南232038

出　　处：《电子技术与软件工程》2024年第4期1-7,共7页ELECTRONIC TECHNOLOGY & SOFTWARE ENGINEERING

基　　金：国家自然科学基金(No.62172183)

摘　　要：针对SAR (Synthetic Aperture Radar)图像在复杂场景下的船舶尺寸小、背景与船舶像素差异度小导致船舶检测精度较低、错检、漏检问题,以YOLOv5s网络为基准模型进行优化,设计跨域多焦点注意力机制(Cross-Domain Multi-focus Mechanism,CDMM)将数据集三维特征转换为二维特征,单独学习通道和空间特征信息并压缩为权值向量,使模型能够在面对复杂背景噪声时保持高度的敏感性和准确性。同时,在基准模型特征金字塔网络(Feature Pyramid Network,FPN)与路径聚合网络(Path Aggregation Network,PAN)特征拼接部分利用自适应注意力机制(Adaptive Attention Mechanism,AAM)引入权值因子,捕捉小目标特征,提高小目标检测能力。实验表明,所提出方法在HRSID数据集上相比基准模型mAP与mAP@0.5:0.95精度提升1.5%与2.4%,与主流模型相比模型优势也十分明显,验证了该方法在提升复杂场景SAR船舶小目标检测效果的有效性。To address the issues of low ship detection accuracy,false positives,and missed detections in SAR(Synthetic Aperture Radar)images in complex scenes due to small ship sizes and small differences between background and shippixels,YOLOv5s network is used as the benchmark model for optimization.A cross domain multi focus mechanism(CDMM)is designed to convert the three-dimensional features of the dataset into two-dimensional features,learn channel and spatial feature information separately,and compress them into weight vectors,enabling the model to maintain high sensitivity and accuracy in the face of complex background noise.At the same time,in the feature pyramid network and path aggregation network of the benchmark model,the feature concatenation part utilizes adaptive attention mechanism(AAM)to introduce weight factors to capture small target features and improve small target detection capability.The experiment shows that the proposed method outperforms the benchmark model on the HRSID dataset mAP and mAP@0.5:0.95 accuracy improvement of 1.5%and 2.4%,and the model has significant advantages compared to mainstream models,which verifies the effectiveness of this method in improving the detection of small ship targets in complex SAR scenes.

关 键 词：SAR图像 YOLOv5s 注意力机制 特征金字塔网络 路径聚合网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]