Lightweight patch-level attention for efficient pig behavior detection: A novel dataset and approach

The pig farming industry plays a crucial role in food production system. Pig behavior detection facilitates timely anomaly detection, enhances farming efficiency, ensures animal welfare, and prevents diseases. Pig behavior detection involves the automatic recognition and classification of pig behaviors in farm images using computer vision and deep learning techniques. Currently, mainstream approaches for pig behavior detection utilize neural networks for image analysis and recognition, however, deep neural network architectures are inherently complex and computationally intensive, with potential for further accuracy enhancement. To address these challenges, we propose a lightweight pig behavior detection model that integrates prediction head optimization and GIoU for improved bounding box regression, while utilizing patch-level attention instead of global attention and employing depthwise separable convolutions in place of standard convolutions. These modifications facilitate efficient and accurate identification and classification of five common pig behaviors in farm environments. Experimental results demonstrate that the proposed method achieves enhanced performance and efficiency.