A Lightweight Posture Recognition Network for Captive Pigs Integrating One-Shot Aggregation Modules
Xiaohong Quan, Deli Zhu, Yu Zhao, Yi Li · 2023
Aiming at the problem that existing pose estimation methods cannot accurately locate the position of key points of pigs in complex captive environments and lead to inaccurate pose recognition, this paper proposes a lightweight network VoV-CenterNet for pig pose recognition. Firstly, the optimized one-shot aggregation network VoVNet39-slim is used as the feature extraction network of the CenterNet, enhancing its feature extraction ability while improving inference speed; Secondly, the Bottleneck module is introduced after ConvTranspose2d convolution layer to solve the problem of insufficient expression ability and overfitting of up-sampling process; Finally, the effectiveness of the attitude recognition method was verified through a self-built dataset. The experimental results show that the recognition accuracy of the VoV-CenterNet network for standing posture is 92.28%, and the recognition accuracy for lying posture is 84.29%. Compared with the benchmark model, the recognition accuracy has increased by 0.75 and 4.11 percentage points respectively, the parameter quantity has decreased by 43.9%, and the inference speed has reached 47FPS.