渔业现代化

渔业现代化 ›› 2023, Vol. 50 ›› Issue (3): 72-78.

• • 上一篇    下一篇

基于通道非降维注意力机制与改进YOLOv5的养殖鱼群检测

  1. (1大连海洋大学 信息工程学院，辽宁 大连 116023；
    2辽宁省海洋信息技术重点实验室，辽宁 大连 116023；
    3设施渔业教育部重点试验室（大连海洋大学），辽宁 大连 116023)
  • 出版日期:2023-06-20 发布日期:2023-06-25
  • 通讯作者: 于红（1968—），女，博士，教授，研究方向：海洋渔业大数据分析、智慧渔业等。E-mail：yuhong@dlou.edu.cn
  • 作者简介:韦思学（1998—），男，硕士研究生，研究方向：计算机视觉。E-mail：546008365@qq.com
  • 基金资助:
    辽宁省重点研发计划项目（2020JH2/10100043）；辽宁省科技重大专项(2020JH1/10200002)；辽宁省教育厅重点科研项目（LJKZ0729）；国家自然科学基金项目（31972846）

A farmed fish detection method based on a non-channel-downscaling attention mechanism and improved YOLOv5#br#
#br#

  1. (1 College of Information Engineering,Dalian 116023, China;
     2 Liaoning Provincial Key of Marine Information Technology,Dalian Ocean University,Dalian 116023, China; 
    3 Key Laboratory of Environment Controlled Aquaculture(Dalian Ocean University), Ministry, 116023, China)

  • Online:2023-06-20 Published:2023-06-25

PDF

摘要： 养殖环境中模糊、气泡遮挡等现象影响养殖鱼特征提取，使养殖鱼群检测精度不佳，为解决上述问题，提出融合通道非降维双重注意力机制ECBAM与改进YOLOv5的养殖鱼群检测模型ESB-YOLO(ECBAM-SPPF-BiFPN-YOLO)。使用ECBAM注意力机制获取更多细节特征；为缓解加入ECBAM导致的检测时间增加、速度变慢，使用SPPF替换SPP，减少模型计算量，降低模型检测时间，提高检测速度；为提高YOLOv5特征融合效果，使用BiFPN进行特征权重融合，提高有效特征在特征融合的比重，减少特征丢失。为了验证改进模块对YOLOv5的影响，设计了消融试验。试验结果显示：ESB-YOLO对比YOLOv5在保持检测速度的条件下平均精度提升了2.40%；设计模型对比试验，验证了ESB-YOLO的优越性，对比FERNet、SWIPENet与SK-YOLOv5等先进水下目标检测模型，ESB-YOLO在平均精度上分别具有3.10%、3.90%与0.70%的优势。研究表明，本研究所提模型对养殖鱼群目标检测效果更佳，可以满足养殖鱼群检测要求。


关键词: 机器视觉, 注意力机制, 养殖鱼检测, YOLOv5, BiFPN

Abstract: The data obtained from the breeding environment has defects such as blurring and occlusion, rendering poor detection of the farmed fish. To solve the above problem, the farmed fish detection model ESB-YOLO (ECBAM-SPPF-BiFPN-YOLO), based on a non-channel-downscaling attention mechanism (ECBAM) and improved YOLOv5, is proposed. ECBAM is adopted to obtain more detailed features. To alleviate the slowdown of detection speed caused by adopting ECBAM, SPPF is used to replace SPP to reduce model computation, decrease model detection time, and improve detection speed. BiFPN is used to perform weighted feature fusion to enhance feature fusion by increasing the proportion of functional features and reducing feature loss. To verify the effectiveness of the proposed modified module for YOLOv5, ablation tests were designed. The test results show that the average precision of ESB-YOLO is improved by 2.40% compared with YOLOv5 while maintaining a similar detection speed. To demonstrate the competitiveness of ESB-YOLO, comparison tests were designed. Compared with advanced underwater object detection models such as FERNet, SWIPENet, and SK-YOLOv5, the average precision of ESB-YOLO is 3.10%, 3.90%, and 0.70% higher, respectively. The study shows that the proposed model is more effective and capable of farmed fish detection.


Key words: machine vision, attention mechanism, farmed fish detection, YOLOv5, BiFPN