Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (18): 190-197.DOI: 10.3778/j.issn.1002-8331.2205-0211

• Graphics and Image Processing • Previous Articles     Next Articles

Improved YOLOv3 Shallow Sea Underwater Biological Target Detection

CHEN Yuliang, DONG Shaojiang, ZHU Sunke, SUN Shizheng, HU Xiaolin   

  1. 1.School of Mechatronics and Vehicle Engineering, Chongqing Jiaotong University, Chongqing 400074, China
    2.Chongqing Industrial Big Data Innovation Center Co., Ltd., Chongqing 404100, China
  • Online:2023-09-15 Published:2023-09-15

改进的YOLOv3浅海水下生物目标检测

陈宇梁,董绍江,朱孙科,孙世政,胡小林   

  1. 1.重庆交通大学 机电与车辆工程学院,重庆 400074
    2.重庆工业大数据创新中心有限公司,重庆 404100

Abstract: Aiming at the low accuracy of shallow underwater biology target detection caused by color distortion, coarse image, local overexposure and large size difference of shallow underwater image, a detection and recognition method of shallow underwater biological target based on YOLOv3 network is proposed. Firstly, the residual attention module is used to solve the problem of local overexposure caused by weak light and artificial light in shallow sea, and improve the feature extraction ability of target. Then, cross-stage local feature extraction is used to improve the detection accuracy, solve the problems of image color distortion and image roughness that affect the detection and recognition effect of shallow underwater biological targets, enhance feature extraction ability and avoid information redundancy. Finally, the calculation function of the total intersection ratio loss is used to solve the problem of large size difference caused by the same species spatial position, which improves the robustness of target detection. The experimental verification proves that the improved YOLOv3 can effectively improve the accuracy and reduce the rate of missed detection in shallow sea underwater biological target detection and recognition task.

Key words: shallow sea underwater biology, target detection, YOLOv3, cross stage partial characteristics, precision

摘要: 针对浅海水下图像存在颜色失真、图像毛糙、局部过曝和尺寸差异大等导致的浅海水下生物目标检测精确度低的问题,提出一种基于YOLOv3网络的浅海水下生物目标检测识别方法。针对浅海中光线弱和使用人造光源导致的局部过曝问题,在YOLOv3网络中加入残差注意力模块,增强对目标的特征提取能力;针对浅海中图像色彩失真和图像毛糙等影响浅海水下生物目标检测识别效果的问题,采用跨阶段局部特征提取提高检测精确度,增强特征提取能力的同时避免信息冗余;针对同一物种由于空间位置造成的尺寸差异大的问题,采用完全交并比损失计算函数,提高目标检测的鲁棒性。通过实验验证证明了改进后的YOLOv3在浅海水下生物目标检测识别任务中能够有效提高精确度,降低漏检率。

关键词: 浅海水下生物, 目标检测, YOLOv3, 跨阶段局部特征, 精确度