Abstract: Aiming at the problems that there are few studies on dead chicken target detection and the large size of the highprecision detection algorithm makes it difficult to deploy to mobile devices, a lightweight dead chicken target detection algorithm based on YOLOv4 is proposed. Firstly, the team collects images of dead chickens in cages from largescale egg production plants to build a target detection dataset. Then, MobileNetv3 backbone extraction network with depthseparable convolution is introduced in the algorithm to reduce the model size. Finally, a selfattentive mechanism module is added before the maximum pooling layer to enhance the algorithms capture of global semantic information. Experimental results in a selfbuilt dataset show that the improved algorithm has higher accuracy in the dead pheasant target detection task, with mAP values and recall rates of 97.74% and 98.15% respectively. The model size is reduced to 1/5 of the original algorithm, and the frame rate reaches 77 frames/s under GPU acceleration, doubling the detection speed and meeting the requirements of embedded deployments.

Key words: identification of dead chicken, deep learning, lightweight network, MobileNet, deep separable convolution

摘要： 针对目前死淘鸡目标检测研究较少，高精度检测算法体积大难以部署至移动式设备等问题，提出一种基于YOLOv4的轻量化死淘鸡目标检测算法。采集大规模蛋鸡养殖工厂笼中死淘鸡图片，建立目标检测数据集；在算法中引入MobileNetv3主干提取网络与深度可分离卷积来降低模型体积；并在最大池化层前添加自注意力机制模块，增强算法对全局语义信息的捕获。在自建数据集中的试验结果表明，改进算法在死淘鸡目标检测任务中有更高的准确度，其mAP值与召回率分别达到97.74%和98.15%，模型大小缩小至原算法的1/5，在GPU加速下帧数达到77帧/s，检测速度提高1倍，能够满足嵌入式部署需求。

关键词: 死淘鸡识别, 深度学习, 轻量化网络, MobileNet, 深度可分离卷积