基于Clite-YOLOv5的鸡状态检测算法
2023,31(9):55-61
摘要：异常鸡的及时发现和处理能够极大地避免规模化养殖中传染病的传播，异常鸡常见的表征状态为鸡闭眼或眯眼。为实现肉鸡养殖过程中问题鸡只的实时监测，提出一种基于YOLOv5的鸡状态检测算法Clite-YOLOv5，该算法以YOLOv5为基础，进行了如下改进：提出融合了CBAM的lite-CBC3模块并用其重构了YOLOv5主干网络，提高了复杂背景下小目标的检测能力；采用改进后的Fuse-NMS抑制算法，降低了检测框的误删率并微调最终检测框；采用深度可分离卷积替换主干网络中的普通卷积，减少了模型的参数量，使模型更易在移动端部署。实验结果表明，提出的Clite-YOLOv5算法均值平均精度（mAP）为92.88%，视频帧率为92FPS，性能超过现有其他算法，能够满足鸡状态实时监测的需求。
关键词：目标检测; 深度学习；鸡状态检测；图像处理；YOLOv5；注意力机制
Clite-YOLOv5 Based Chicken State Detection Algorithm
Abstract：The timely detection and treatment of abnormal chickens can greatly avoid the spread of infectious diseases in large-scale farming, and the common characterization state of abnormal chickens is that they close their eyes or squint. To achieve real-time monitoring of problem chickens in broiler farming, a YOLOv5-based chicken status detection algorithm Clite-YOLOv5 is proposed, which is based on YOLOv5 with the following improvements: the lite-CBC3 module incorporating CBAM is proposed and used to reconstruct the YOLOv5 backbone network to improve the detection capability of small targets in complex backgrounds; The improved Fuse-NMS suppression algorithm is used to reduce the false deletion rate of detection frames and fine-tune the final detection frames; the depth-separable convolution is used to replace the normal convolution in the backbone network, which reduces the number of parameters of the model and makes the model easier to deploy on mobile. Experimental results show that the proposed Clite-YOLOv5 algorithm has a mean average precision (mAP) of 92.88% and a video frame rate of 92 FPS, which outperforms other existing algorithms and can meet the demand for real-time monitoring of chicken status.
Key words：Target detection; Deep learning; Chicken state detection; Image processing; YOLOv5; Attention mechanism
收稿日期：2023-03-20
基金项目：国家自然科学基金(62003183)。
