Abstract: The intelligent feeding strategy is the key to efficiently using bait and significantly reducing breeding costs in industrial recirculating aquaculture. Therefore, this study proposes an intelligent feeding strategy that combines texture discrimination of water surface images and residual baits detection based on YOLOv5-BCH. Firstly, taking the calm water surface as the baseline, the residual baits recognition frame is obtained through the texture features of the water surface image during the feeding process. The BottleNet-CSP module is employed to enhance the Backbone, and the CBAM module is utilized in the Neck, improving spatial and channel dimensions. This integration effectively fuses multi-scale features. Furthermore, to enhance the detection accuracy further, we introduce three micro-scale detection heads in the Head section. This configuration enhances the network's capacity to detect small targets on the water surface, resulting in significant improvements in mAP0.5 (40.26%), mAP0.5:0.95 (15.59%), and Precision (37.85%). The adaptive feeding strategy of "trial feeding+single wheel multiple times" is adopted, effectively reducing labor input and feed waste. The results show that this system can replace manual feeding and achieve intelligent feeding throughout the process, providing a valuable reference for realizing unmanned feeding in industrial aquaculture.

Key words: intelligent feeding strategy, machine vision, deep learning, residual baits recognition, recirculating aquaculture.
 

摘要： 智能投喂策略是实现工厂化循环水养殖过程中饲料高效利用、降低养殖成本的关键。提出一种结合水面图像纹理判别和YOLOv5-BCH残饲检测的智能投喂策略。首先，以平静水面为基线，通过摄食过程水面图像纹理特征获得残饲识别帧；其次，通过采用BottleNet-CSP模块与CBAM模块分别对YOLOv5的Backbone和Neck端进行改进，增强了深度神经网络在空间和通道维度上的特征表示能力，且有效融合了多尺度特征。同时在Head部分设置3个微尺度检测头增强对水面小目标特征的捕捉能力，使mAP0.5、mAP0.5:0.95和精确率分别提升40.26%、15.59%和37.85%；最后，设计智能投喂系统并采用“试投+单轮多次”自适应投喂策略，有效降低了劳动力投入及饲料浪费。研究表明，该系统可代替人工实现全流程智能化投喂，为工厂养殖饲喂环节实现无人化提供参考。

关键词: 智能投饲策略, 机器视觉, 深度学习, 残饲识别, 工厂化循环水养殖