Sow Behavior Object Detection Network Using Swin Transformer

Abstract

Observing sow behavior constitutes a critical aspect of sow farming management. It allows breeders to promptly assess the need for intervention in the piglet production process, ultimately playing a pivotal role in enhancing piglet survival rates and overall farm profitability. This research introduces a sow behavior detection approach founded on Swin Transformer. Initially, we gathered and annotated images of sows within the pigpen, creating a dataset for sow posture detection. Leveraging this dataset, we trained a sow behavior detection model based on Swin Transformer. After testing, the Swin Transformer sow behavior detection model mAP (mean Average Precision) trained in the article reached a maximum of 0.85. The time required for single sow behavior recognition and inference was about 8ms, and behavior detection and inference was performed once every 200 consecutive frames of posture extraction, Update the behavior detection results once every 23.3 seconds on average. Prove the feasibility of the proposed Swin Transformer based sow pose detection method, providing ideas for sow pose detection in large farms.