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

Area of Science:

• Agricultural Science
• Computer Vision
• Animal Science

Background:

• Traditional vision-based pig behavior recognition methods struggle with feature robustness and spatial-channel modeling in complex environments.
• Limitations in current methods hinder reliable performance in automated animal health monitoring systems.

Purpose of the Study:

• To propose a novel Dual-path Swin Transformer (DST) framework to address limitations in pig behavior recognition.
• To enhance the accuracy and robustness of pig behavior recognition in challenging agricultural settings.

Main Methods:

• Developed a Dual-path Swin Transformer (DST) framework with frequency-domain and spatial-domain paths.
• Introduced a Frequency-domain Fusion Filter Module (FFM) using Ideal Low-pass and Gaussian High-pass Filters.
• Implemented a Decoupled Spatial-Channel Attention (DSCA) mechanism for enhanced feature learning.

Main Results:

• The DST framework achieved a pig behavior recognition accuracy of 94.94% on a dataset of 2,755 video clips.
• DST outperformed the baseline Swin Transformer by 1.45 percentage points.
• The proposed methods demonstrated improved feature robustness and spatial-channel dependency modeling.

Conclusions:

• The Dual-path Swin Transformer (DST) offers an effective and robust solution for automated pig behavior monitoring.
• The DST framework shows significant potential for improving animal health monitoring in complex agricultural environments.
• Enhanced feature learning through frequency and spatial domain processing is key to reliable behavior recognition.