Enhancing Abnormal Behavior Recognition in Group-Housed Pigs Through the MS-FA-DETR: A Multi-Scale and Frequency-Aware Fusion Method

In contemporary pig farming, monitoring pig behavior is crucial as it reflects their health and well-being. Existing behavior recognition methods struggle with abnormal behaviors and complex environments, including challenges in distinguishing between small or marginal animal movements and separating behaviors from background noise. This study presents MS-FA-DETR, a novel deep learning model that significantly improves the detection and analysis of pig behaviors in challenging farm settings. The model features the Enhanced Scale-Selective Channel Attention (ES-SCSA) mechanism, which substantially boosts the extraction of multi-scale features, adeptly addressing detection challenges related to small and marginal targets. Additionally, it incorporates a Frequency Bi-directional Feature Pyramid Network (FreqBiFPN) that expertly separates pig behaviors from background through frequency-aware processing, enhancing its discriminative capabilities. Furthermore, the Inner-MPD-IoU loss function is customized to optimize the model's ability to accurately define target relationships in complex settings, thereby improving localization accuracy. A comprehensive dataset, PigABRD, comprising three abnormal behaviors (fighting, mounting, climbing fences) and four common behaviors (sniffing, walking, lying, kneeling) was established. Experimental results demonstrate that the precision, recall, mAP50, and mAP50-95 of our model are respectively <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$93.8 \%, 92.4 \%, 96.7 \%$</tex>, and 84.6 %, showing remarkable improvements over the RT-DETR model and outperforming established models such as Faster RCNN, CenterNet, EfficientDet, and the YOLO series. MS-FA-DETR provides vital technical support for intelligent farming initiatives and promotes pig welfare.