Abstract: Feeding is one of the main tasks in aquaculture, and how to reduce the cost of feeding is the key point of maximizing the profits of aquaculture. Using computer vision technology to monitor the feeding behavior of fish and quantify the intensity of fish feeding desire can realize automatic feeding on demand, reduce feed waste, and improve feed utilization. This paper reviews the research progress of four kinds of fish intelligent feeding methods based on computer vision, such as food detection, optical flow method, texture and deep learning, and analyzes the advantages and disadvantages of each method. Bait detection method is simple and easy to implement, but it is difficult to identify the residual bait accurately. Optical flow method can capture fish movement information effectively, but it is easily affected by environment and light. Texture and other feature methods use more types of features and more effective information, but they are not suitable for high-density culture. Deep learning method has high recognition accuracy, strong robustness, large calculation amount, and high requirements for equipment computing power. Based on this, three research directions of large-scale data set, efficient lightweight depth model and “Internet of Things +” intelligent feeding are proposed. It provides reference for further improving the maturity and practicability of intelligent feeding method.

Key words: fish, computer vision, feeding behavior, feed detection, optical flow, deep learning, intelligent feeding

摘要： 饵料投喂是水产养殖中的主要工作之一，降低饵料投饲成本是水产养殖利润最大化的关键。利用计算机视觉技术对鱼类摄食行为进行监测，量化鱼类摄食欲望强度，可以实现自动按需投饵，减少饵料浪费，提高饵料利用率。综述饵料检测法、光流法、纹理特征法和深度学习这4种基于计算机视觉的鱼类智能投喂方法的研究进展，对各方法的优缺点进行深入分析。饵料检测法简单易实现，但难以准确识别残饵；光流法能有效捕捉鱼群运动信息，但易受环境和光照影响；纹理等特征法使用的特征种类更多，有效信息更丰富，却不适用于高密度养殖；深度学习法识别精度高、鲁棒性强，但计算量大，对设备算力要求高。基于此，提出建立大规模数据集、构建高效轻量级深度模型和“物联网+”智能投喂3个研究方向，为进一步提升智能投喂方法的成熟度和实用性提供参考。

关键词: 鱼类, 计算机视觉, 摄食行为, 饵料检测, 光流法, 深度学习, 智能投喂