基于SEEC-YOLO v5s的散养蛋鸡日常行为识别与统计系统

针对鸡只个体较小、个体间存在遮挡，对蛋鸡日常行为识别造成干扰的问题，提出了一种基于SEEC-YOLO v5s的蛋鸡日常行为识别方法。通过在YOLO v5s模型输出部分添加SEAM注意力模块、在特征融合部分引入显式视觉中心模块(EVCBlock),扩大了模型的感受野，提高了模型对小个体遮挡情况下的目标识别能力，提升了模型对蛋鸡站立、采食、饮水、探索、啄羽和梳羽6种行为的识别精度。提出了一种基于视频帧数与视频帧率比值计算蛋鸡日常行为持续时间的统计方法，并对蛋鸡群体一天之中不同时间段及全天各行为变化规律进行了分析。将改进后的模型进行封装、打包，设计了蛋鸡日常行为智能识别与统计系统。试验结果表明，SEEC-YOLO v5s模型对6种行为识别的平均精度均值为84.65%,比YOLO v5s模型高2.34个百分点，对比Faster R-CNN、YOLO X-s、YOLO v4-tiny和YOLO v7-tiny模型，平均精度均值分别提高4.30、3.06、7.11、2.99个百分点。本文方法对蛋鸡的日常行为监测及健康状况分析提供了有效的支持，为智慧养殖提供了借鉴。

Daily Behavior Recognition and Real-time Statistics System of Free-range Laying Hens Based on SEEC-YOLO v5s

The small size of the chickens and the shading of the chickens from each other are factors that make it difficult to identify the daily behaviour of laying hens. To address this problem, a method of daily behavior identification of laying hens based on SEEC-YOLO v5s was proposed. By adding a SEAM attention module (separated and enhancement attention module) to the output part of the YOLO v5s model and introducing an EVCBlock module (explicit visual center) to the feature fusion part, the perceptual field of the model was expanded, the recognition ability of the model for occluded targets was improved, and the recognition accuracy of the model for the six behaviors of standing, feeding, drinking, exploring, feather pecking and grooming of laying hens was improved. A statistical method was proposed to calculate the duration of daily behavior of laying hens based on the ratio of video frames to video frame rate, and various behavioral changes of laying hens at different times of the day and throughout the day were analyzed. The improved model was encapsulated and packaged to develop an intelligent identification and automatic statistics system for the daily behavior of laying hens. The test results showed that the mAP of SEEC-YOLO v5s model for six behaviors recognition was 84.65%, which was 2.34 percentage points higher than that of YOLO v5s model, and compared with that of Faster R-CNN, YOLO X-s, YOLO v4-tiny and YOLO v7-tiny models, the mAP was improved by 4.30 percentage points, 3.06 percentage points, 7.11 percentage points and 2.99 percentage points, respectively. The method can provide effective support for daily behavior monitoring and health condition analysis of laying hens, and provide a reference for smart farming.