摘要
舰船已经成为海上军事重要的监测目标,针对SAR图像舰船目标检测存在检测效果差、计算量大、泛化能力弱的问题,提出一种基于YOLOv5和Mobilenetv3的轻量化舰船目标检测算法。首先,引入Mobilenetv3主干网络,降低模型计算量与体积,实现模型轻量化处理;然后,引入EIoU损失函数,提高预测框的回归精度和收敛速度;最后,在颈部网络中引入CBAM,在特征融合阶段进行注意力调整,提高模型检测精度与检测效果。在SSDD数据集上的实验结果显示,改进后算法模型体积压缩至原YOLOv5模型的18.32%,训练时间缩短35.22%,参数量减小至原模型的15.94%,计算量减小至原模型的10.76%,平均精度提升至98.3%。实验结果表明,改进后算法在保持高精度检测效果的情况下,大幅降低了参数量和计算量,减小了模型体积,并缩短了训练时间。
The ship has become an important monitoring target in maritime military field.Ship target detection in SAR images suffers from poor detection effects,large computation amount and weak generalization capability.To solve the problems,a lightweight ship target detection algorithm based on YOLOv5 and Mobilenetv3 is proposed.Firstly,Mobilenetv3 backbone network is introduced to reduce the computation amount and volume of the model and realize lightweight processing of the model.Then,the EIoU loss function is introduced to improve the regression accuracy and convergence speed of the prediction box.Finally,CBAM is introduced into the neck network,and attention adjustment is conducted at the stage of feature fusion to improve the detection accuracy and detection effects of the model.The experimental results on SSDD dataset show that the volume of the improved algorithm model is reduced to 18.32%of that of the original YOLOv5 model,the training time is shortened by 35.22%,the parameter quantity is reduced to 15.94%of that of the original model,the computation amount is reduced to 10.76%of that of the original model,and mAP is improved to 98.3%.The experimental results show that the improved algorithm greatly reduces parameter quantity,computation amount,model volume and training time while maintaining high-precision detection effects,which can realize real-time detection of ship targets in SAR images.
作者
张上
陈益方
王申涛
王恒涛
冉秀康
ZHANG Shang;CHEN Yifang;WANG Shentao;WANG Hengtao;RAN Xiukang(China Three Gorges University,College of Electrical Engineering and New Energy,Yichang 443000,China;China Three Gorges University,Hubei Province Engineering Technology Research Center for Construction Quality Testing Equipment,Yichang 443000,China;China Three Gorges University,College of Computer and Information,Yichang 443000,China;Nantong Institute of Technology,Nantong 226000,China)
出处
《电光与控制》
CSCD
北大核心
2023年第12期66-72,共7页
Electronics Optics & Control
基金
国家级大学生创新创业训练计划(202011075013)。