摘要
针对舰船及要害部位检测任务的需求,设计了SEC-YOLOv5(Swin Transformer C3EE CBAM-YOLOv5)舰船及要害部位检测算法。SEC-YOLOv5算法利用Swin Transformer改进网络主干部分C3模块,增强模型对语义特征的提取能力,提升要害部位小目标的识别精度;使用CBAM注意力机制加权有效特征信息,提升特征信息利用率;设计C3EE(C3 easy effect)模块替换瓶颈层所有C3模块,扩大模型感受野,丰富模型的梯度信息流。基于RS-ship数据集验证所提算法,以原YOLOv5算法为基准进行逐模块的消融对比实验,实验结果表明,SEC-YOLOv5算法在增加少量参数量的前提下,将平均精度值mAP@0.5提高5个百分点。
According to the requirements of ships and key parts,this paper designs the detection algorithm of SEC-YOLOv5 ships and key parts.The SEC-YOLOv5 algorithm uses Swin Transformer to improve the C3 module of the backbone network,enhancing the ability to extract semantic features and improving the identification accuracy of small targets in key parts.It also uses the CBAM attention mechanism to weight effective feature information and improve the utilization rate of feature information.The paper designs the C3EE module to replace all C3 modules in the bottleneck layer,expand the feeling field of the model,and enrich the gradient information flow of the model.Based on the RS-ship data set,the proposed algorithm is verified,and the original YOLOv5 algorithm is used as the benchmark for the module ablation comparison experiment.The experimental results show that the SEC-YOLOv5 algorithm increases the average accuracy value mAP@0.5 by 5 percentage points on the premise of increasing a small number of parameters.
作者
高兵
祝宇鸿
GAO Bing;ZHU Yuhong(College of Communication Engineering,Jilin University,Changchun 130000,China)
出处
《计算机工程与应用》
北大核心
2025年第11期284-294,共11页
Computer Engineering and Applications
