Abstract: In order to realize the rapid and accurate identification of surface diseases of fish infected by viruses and bacteria in intensive aquaculture, and help farmers quickly understand the damage degree and distribution of fish diseases in the aquaculture tank, a set of rapid detection system for fish diseases was designed based on the improved YOLOv5 combined with embedded technology. The improved YOLOv5 neural network model was used to generate the candidate boxes of fish diseases and realize the rapid classification of fish diseases. The detection system counted and classified fish diseases according to the data in the candidate box. The classification of fish disease hazards was divided into normal, mild and severe, and a quantitative evaluation system for the degree of fish disease hazards was formed by combining the number of fish with the number of sick fish. Finally, GPRS module was introduced to obtain the location information of detection points and form a heat map of fish diseases at the software end. Model test results showed that the detection accuracy of the improved YOLOv5 model was 99.75%, and the recall rate was 93.21%. Test models mAP50 and mAP50: 95 compared with the original YOLOv5 model, the AP reached 99.38% and 88.09% in the case of a slight decrease in FPS of 3.22 frames, indicating excellent performance, and the memory of the improved model was reduced to 13.6 MB. The improved YOLOv5 model had smaller size, superior performance and strong stability, and was suitable for deployment in the embedded system of fish disease detection. The overall test results of the system show that the system can detect the occurrence of fish disease in real time, and the system can classify the fish disease according to normal, mild and severe, and combine the situation of fish disease with the positioning system to form a visual thermal image.

Key words: fish disease detection, YOLOv5, picture processing, feature extraction, information service

摘要： 为实现集约化水产养殖中的鱼类因病毒细菌等感染体表病症的快速、准确识别，帮助养殖户快速了解养殖池内的鱼病危害程度和分布情况，基于改进的YOLOv5结合嵌入式技术设计一套鱼病的快速检测系统。使用改进过的YOLOv5神经网络模型生成鱼病的候选框，实现对鱼病的快速定级分类。检测系统根据候选框的数据对鱼病进行计数、分类，鱼病危害分类按正常、轻度、重度划分，结合患病鱼数形成对鱼病危害程度定量化测评的体系，最后引入GPRS模块获取检测点的位置信息，在软件端形成鱼病的热力图。模型测试结果表明：改进后的YOLOv5模型检测精准率为99.75％，召回率为93.21%，测试模型mAP50、mAP50：95对比原YOLOv5模型在帧数轻微下降3.22帧的情况下AP达到99.38%、88.09%，表明其拥有出色性能，改进后模型内存下降至13.6 MB。改进后YOLOv5模型体积更小，性能优越稳定强，适宜部署在鱼病检测嵌入式系统中。系统整体测试结果表明：系统能够实时的检测鱼病的发生，检测时系统能按正常，轻度，重度划分鱼病，并将鱼病的情况结合定位系统形成可视化的热力图像。

关键词: 鱼病检测, YOLOv5, 图像处理, 特征提取, 信息服务