Authors

Abstract

Temperature is a very important factor in the breeding of yellow feather broilers. Researching the group behaviour of yellow feather broilers under heat stress can help farmers take the corresponding measures to reduce the heat stress of broilers and improve production performance. In this paper, several traditional methods have been employed to detect and locate the broilers. These methods are highly interfered by the background whose colour is similar to broilers, thereby making it difficult to accurately locate the broilers. Meanwhile, although the algorithm YOLOv3 can precisely segment and locate the broilers, it has the disadvantages of incomplete detection and low detection confidence rate. Finally, this paper applies the neural architecture search and transfer learning to train the pre-processed training set, and obtains a detection model with a recognition accuracy of 83%. Then this model is used to process the images under heat stress at every 30 s so as to obtain the distribution of the broilers at each moment. Based on the results, the liveness and distribution characteristics of the broilers are analyzed. The analysis results show that when heat stress occurs, the broilers mainly gather at the vent; that when the temperature further rises above 30°C, the proportion of broilers at the vent increases from 53.3% to 67.3% on average; and that the activity index of the broilers decreases by 22.54% within an average of 3 h after the temperature rises to 30°C.

Abstract in Chinese

在黄羽鸡生产养殖中，温度是一个非常重要的因素。研究热应激下黄羽鸡的群体行为可以帮助养殖户更深入地了解黄羽鸡在热应激下的反应，从而采取相应的措施减少鸡的热应激现象，提高鸡的生产性能。本文首先研究了几种对黄羽鸡检测的方法，根据检测的结果无法进行下一步的精确分析，而基于Darknet-53的YOLOv3算法则存在识别不精准的问题。最后本文利用Neural Architecture Search 和Transfer Learning对预处理后的训练集进行训练，获得了识别准确率为83%的检测模型。基于此模型分析鸡群活跃度和分布特征，结果表明，鸡群在温度上升至30℃之后平均3小时内活跃度指数会降低22.54%。在分布上，热应激发生时鸡只主要聚集在通风口处，当温度进一步升高至30℃以上时，通风口处鸡的比例平均由53.3%上升至67.3%