基于改进YOLO v6-tiny的蛋鸡啄羽行为识别与个体分类

针对目前蛋鸡啄羽异常行为（包括啄和被啄）识别精度比较低的问题,提出了一种基于改进YOLO v6-tiny模型进行啄羽异常行为识别的方法。该方法通过在YOLO v6-tiny模型中引入DenseBlock结构并融入CSP结构的SPP模块（SPPCSPC）的方式,增强了YOLO v6-tiny模型的特征提取能力,扩大了模型的感受野,提升了模型的检测精度。在识别出啄羽异常行为的基础上,对如何基于异常行为发生次数,进行蛋鸡个体分类进行了研究。提出了基于YOLO v6-tiny模型进行蛋鸡个体识别,并将啄羽异常行为识别结果输入个体识别网络,进行蛋鸡个体分类的方法。同时,本文还分别对2种不同的养殖密度、一天当中3个不同的时间段,异常行为发生次数的变化规律进行了分析。实验结果表明,优化后的模型对啄和被啄异常行为的识别平均精度（AP）分别为92.86%和92.93%,分别比YOLO v6-tiny模型高1.61、1.08个百分点,比Faster R-CNN模型高3.28、4.00个百分点,比YOLO v4-tiny模型高6.15、6.63个百分点,比YOLO v5s模型高2.04、4.27个百分点,比YOLO v7-tiny模型高5.39、3.92个百分点。本文方法可以识别出啄和被啄羽异常行为,为蛋鸡异常行为的智能检测提供了技术支撑。

Feather Pecking Abnormal Behavior Identification and Individual Classification Method of Laying Hens Based on Improved YOLO v6-tiny

To address the current problem of low accuracy in the recognition of feather pecking anomalies (including pecking and pecked) in laying hens, a method for feather pecking anomaly recognition was proposed based on an improved YOLO v6-tiny model. By introducing the DenseBlock structure into the YOLO v6-tiny model and incorporating the SPP module SPPCSPC into the CSP structure, the feature extraction capability of the YOLO v6-tiny model was enhanced, the sensory field of the model was expanded, and the detection accuracy of the model was improved. Based on the identification of feather pecking anomalies, how to classify individual laying hens was investigated based on the number of anomalies. The method to identify individual laying hens based on the YOLO v6-tiny model was proposed and identification results of feather pecking anomalies were input into the individual identification network to classify individual laying hens. At the same time, the change pattern of the number of anomalies at two different breeding densities and three different times of the day was also analyzed. The experimental results showed that the average precision (AP) of the optimized model were 92.86% and 92.93% for pecking and pecked anomalies, respectively, which were 1.61 percentage points and 1.08 percentage points higher than that of the YOLO v6-tiny model, 3.28 percentage points and 4.00 percentage points higher than that of the Faster R-CNN model, 6.15 percentage points and 6.63 percentage points higher than that of the YOLO v4-tiny model, 2.04 percentage points and 4.27 percentage points higher than that of the YOLO v5s model, and 5.39 percentage points and 3.92 percentage points higher than that of the YOLO v7-tiny model. The method can identify the abnormalities of pecking and pecked feathers, which provided technical support for the intelligent detection of abnormal behavior of laying hens. The results of classifying individual laying hens based on pecking abnormalities provided a basis for preferential breeding of individual laying hens.