基于改进YOLO和迁移学习的水下鱼类目标实时检测  

Real-Time Detection of Underwater Fish Based on Improved YOLO and Transfer Learning

作　　者：李庆忠[1] 李宜兵 牛炯[1] LI Qingzhong;LI Yibing;NIU Jiong(College of Engineering, Ocean University of China, Qingdao 266100)

机构地区：[1]中国海洋大学工程学院,青岛266100

出　　处：《模式识别与人工智能》2019年第3期193-203,共11页Pattern Recognition and Artificial Intelligence

基　　金：国家重点研发计划项目(No.2017YFC1405202);国家自然科学基金项目(No.61132005);海洋公益性行业科研专项(No.201605002)~~

摘　　要：为了实现非限制环境中水下机器人基于视频图像的水下鱼类目标快速检测,提出基于改进YOLO和迁移学习的水下鱼类目标实时检测算法.针对YOLO网络的不足,设计适合水下机器人嵌入式系统计算能力的精简YOLO网络(Underwater-YOLO).利用迁移学习方法训练Underwater-YOLO网络,克服海底鱼类已知样本集有限的限制.利用基于限制对比度自适应直方图均衡化的水下图像增强预处理算法,克服水下图像的降质问题.利用基于帧间图像结构相似度的选择性网络前向计算策略,提高视频帧检测速率.实验表明,文中算法能实现在非限制环境下海底鱼类目标的实时检测.相比YOLO,文中算法对海底鱼类小目标和重叠目标具有更好的检测性能.To fast detect underwater fish in unrestricted underwater environment based on underwater video collected by underwater robots, a real-time detection algorithm for underwater fish based on improved you only look once(YOLO) and transfer learning is proposed. Firstly, an underwater-YOLO for the embedding computer system of underwater robots is designed to overcome the shortcomings of traditional YOLO. Then, transfer learning strategy is employed to train the underwater-YOLO network and alleviate the limitation of known underwater fish samples. A preprocessing algorithm based on contrast limited adaptive histogram equalization is proposed to overcome the problem of underwater image degradation. Finally, a video frame selection method for foreground computation of underwater-YOLO based on structure similarity between inter-frames is proposed to increase the detection frame rate. The experimental results show that the proposed algorithm achieves the goal of real-time detection of underwater fish in unconstrained underwater environment. Compared with the traditional YOLO, the proposed underwater-YOLO generates better detection performance in complex scenes with small fish and overlapped fishes.

关 键 词：卷积神经网络 深度学习 鱼类目标检测 单级式目标检测算法(YOLO) 迁移学习 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]