文献类型：会议论文

英文题名：Reserve sow pose recognition based on improved YOLOv4

作者：Lu, Lianfeng[1]; Mao, Liang[2]; Wang, Ji[1]; Gong, Wenchao[1]

机构：[1] Guangdong Ocean University, College of Electronic and Information Engineering, Zhanjiang, China; [2] Institute of Artificial Intelligence Application Technology, Guangdong-Hong Kong-Macao Greater Bay Area Shenzhen Polytechnic, Shenzhen, China

会议论文集：2022 IEEE Asia-Pacific Conference on Image Processing, Electronics and Computers, IPEC 2022

会议日期：April 14, 2022 - April 16, 2022

会议地点：Dalian, China

语种：英文

外文关键词：ESCMobileNetv3-YOLOv4; lightweight; Posture recognition; Reserve sow

外文摘要：To address the problems of low accuracy of reserve sow pose recognition and high computational load of the algorithm, a reserve sow pose recognition method based on ESCMobileNetv3-YOLOv4 is proposed in this paper. Firstly, in order to solve the problem of over-fitting caused by unbalanced samples, Few Label Guide Data Augmentation (FLGDA) technique is designed for data augmentation. Secondly, with YOLOv4 as the basic network model, ESCMobileNetv3 is obtained by introducing Extremely Separated Convol in MobileNetv3, which is used as the backbone network of YOLOv4 to improve the recognition accuracy and reduce the number of parameters. Finally, Embedding depth separable convolution in feature fusion and output prediction makes the model further lightweight. Experiments showed that the average recognition accuracy of the lightweight model ESCMobileNetv3-YOLOv4 was 97.44%, 96.43%, 96.57% and 96.89% for the four types of postures of standing, sitting, ventral recumbent and lateral recumbent of reserve sow posture definitions, respectively, and the average recall rate was 95.67%, with the average accuracy mean value reaching 96.83%, which was better than the original The model parametric number is 86.1% smaller than YOLOv4, which saves memory space and facilitates portability to removable devices. The single-frame inference speed on Nvida RTX 2070 is 43.8 % higher than YOLOv4, which improves the efficiency of sow pose recognition. ? 2022 IEEE.