摘要
海面舰船目标检测容易受陆地、海浪等背景的干扰。针对舰船小目标检测精度低和鲁棒性差的问题,提出一种改进的舰船目标检测模型CWMA-YOLOv5s。首先,设计具有多分支跨层连接的C2f模块丰富多目标舰船梯度流信息。然后,设计并实现了残差多头自注意力融合模块优化特征融合效果。其次,改进Predection网络,设计SCP结构,提高了舰船目标的显著度。最后,引入改进的WIOU损失函数解决CIOU损失函数带来的梯度爆炸和模型提前退化问题。实验结果表明,与YOLOv5s模型相比,该模型在MASATI-v2数据集上,精度提高了13.1%,召回率提高了12.8%,mAP@50提高了6.8%。与其他同类型检测算法相比,该算法拥有更好的学习能力,整体检测精度达到了82.3%,具有较强的鲁棒性。
The detection of ship target on sea surface is easy to be interfered by the background such as land and sea wave.Aiming at the problems of low precision and poor robustness of ship small target detection,an improved ship target detection model CWMA-YOLOv5s is proposed.Firstly,a C2f module with multi-branch cross-layer connections is designed to enrich multi-target ship gradient flow information.Then,a residual polytope self-attention fusion module is designed and implemented to optimize the feature fusion effect.The Predection network is then improved and the SCP structure is designed to increase the saliency of the ship's target.Finally,an improved WIOU loss function is introduced to solve the problems of gradient explosion and early model degradation caused by the CIOU loss function.The experimental results show that the model improves precision by 13.1%,improves recall by 12.8% and improves mAP@50 by 6.8%on the MASATI-v2 dataset compared to the YOLOv5s model.Compared with other detection algorithms of the same type,the algorithm has better learning ability,the overall detection accuracy reaches 82.3%,and has strong robustness.
作者
师红宇
蔡自桂
杜文
张哲于
SHI Hongyu;CAI Zigui;DU Wen;ZHANG Zheyu(School of Computer Science,Xi'an Polytechnic University,Xi'an 710048)
出处
《舰船电子工程》
2025年第2期34-38,73,共6页
Ship Electronic Engineering
基金
陕西省重点研发计划项目(编号:2022GY-058)
西安市科技创新人才服务企业项目(编号:2020KJRC0022)资助。
作者简介
师红宇,女,高级工程师,研究方向:图像处理、深度学习、智能检测;蔡自桂,男,硕士研究生,研究方向:遥感图像处理、目标检测;杜文,女,硕士研究生,研究方向:密集人群估计、目标检测;张哲于,男,硕士研究生,研究方向:深度学习、目标检测。
