文章摘要
孟俊霞,王新亮.基于YOLOv8-DCW的水下鱼类目标检测方法[J].安徽建筑大学学报,2025,33(5):56-64
基于YOLOv8-DCW的水下鱼类目标检测方法
Underwater Fish Target Detection Method Based on YOLOv8-DCW
  
DOI:
中文关键词: 水下鱼类目标检测  YOLOv8  可变形卷积  WIOU损失
英文关键词: fish target detection  YOLOv8  deformable convolution  WIOU loss
基金项目:国家自然科学基金项目(41906168);安徽省自然科学基金项目(2308085MD124);安徽省高校自然科学研究项目(2022AH050256);安徽省高校协同创新项目(GXXT-2022-020);安徽建筑大学结余经费资助项目(JZ202366)
作者单位
孟俊霞 College of Civil EngineeringAnhui Jianzhu UniversityHefei 230601China 
王新亮 College of Civil EngineeringAnhui Jianzhu UniversityHefei 230601China 
摘要点击次数: 194
全文下载次数: 0
中文摘要:
      为了解决水下鱼类目标检测中因环境复杂所致的精度不足的问题,提出一种基于YOLOv8-DCW模型的检测方法。首先,通过随机旋转+随机缩放、随机裁剪+对比度增强两种方式扩充水下鱼类光学图像数据集;然后,利用分层级可变形卷积改进YOLOv8骨干网络深层、颈部网络浅层、头部网络浅层结构,以提高模型对细节特征捕捉、特征增强、特征融合的能力;最后,引入具有动态聚焦机制和简化惩罚项计算的损失函数 WIOU(Wise-IOU)解决CIOU(Complete Intersection over Union)函数对水下小目标鱼类、被遮挡目标以及处理低质量样本的局限性,以提高模型计算速度和检测精度。实验结果表明,YOLOv8-DCW 模型的 mAP0.5 值达到 94.9%,较原模型提升 4.5%;mAP0.5:0.95值达到77.2%,较原模型提升8.1%,有效提升水下鱼类目标的检测精度,漏检情况也有所改善。
英文摘要:
      To solve the problem of insufficient accuracy in underwater fish target detection caused by complex environments, a detection method based on YOLOv8-DCW model is proposed. Firstly, the underwater fish optical image dataset is expanded through two methods: random rotation+random scaling, and random cropping+contrast enhancement; Then, the hierarchical deformable convolution is used to improve the deep, neck, and head structures of the YOLOv8 backbone network, in order to enhance the model′s ability to capture, enhance, and fuse detailed features; Finally, a loss function WIOU(Wise OU)with dynamic focusing mechanism and simplified penalty calculation is introduced to address the limitations of CIOU(Complete Intersection over Union)function for underwater small target fish, occluded targets, and processing low-quality samples, in order to improve the model calculation speed and detection accuracy. The experimental results showed that the mAP0.5 value of the YOLOv8-DCW model reached 94.9%, an increase of 4.5% compared to the original model. The mAP0.5:0.95 value reached 77.2%, an increase of 8.1% compared to the original model.Which effec⁃ tively improved the detection accuracy of underwater fish targets.
查看全文   查看/发表评论  下载PDF阅读器
关闭