DST: a Dual-path Swin Transformer framework for pig behavior recognition

Pig behavior recognition is a crucial component of intelligent animal health monitoring. In complex pigpen environments, traditional vision-based methods often exhibit poor feature robustness and deficient spatial-channel dependency modeling, two key limitations that compromise reliable performance. To mitigate these limitations, this study proposes a Dual-path Swin Transformer (DST) framework. This framework consists of two complementary paths for feature learning: a frequency-domain path and a spatial-domain path with enhanced channel modeling. In the frequency-domain path, a novel Frequency-domain Fusion Filter Module (FFM) is introduced. The Ideal Low-pass Filter extracts coarse-scale global structural features, while the Gaussian High-pass Filter captures fine-grained local edge features. In the spatial-domain path, an effective Decoupled Spatial-Channel Attention (DSCA) mechanism is developed. The spatial attention branch adaptively enhances features in key regions, and the channel attention branch automatically strengthens the weights of feature channels highly correlated with pig behaviors. The proposed DST is validated on a dataset containing 2,755 video clips covering six typical behaviors. Results show that DST achieves a recognition accuracy of 94.94%, which is 1.45 percentage points higher than the baseline Swin Transformer. These findings demonstrate that DST provides an effective and robust solution for automated pig behavior monitoring in complex agricultural environments.