Automatic Detection and Segmentation for Group-Housed Pigs Based on PigMS R-CNN.

Area of Science:

• Computer Vision
• Animal Science
• Agricultural Technology

Background:

• Accurate segmentation of individual pigs in group housing is crucial for monitoring health and welfare.
• Existing methods struggle with identifying pigs in close proximity or overlapping areas.

Purpose of the Study:

• To develop and evaluate an instance segmentation framework, PigMS R-CNN, for precise identification and localization of pigs in group-housed environments.
• To improve upon traditional non-maximum suppression (NMS) techniques for better detection accuracy in challenging scenarios.

Main Methods:

• The PigMS R-CNN framework utilizes a ResNet-101 and Feature Pyramid Network (FPN) for feature extraction.
• Region candidate network generates regions of interest (RoIs), followed by regression, classification, and mask prediction branches.
• Soft non-maximum suppression (soft-NMS) was implemented to replace traditional NMS for improved post-processing of detected pigs.

Main Results:

• The PigMS R-CNN framework achieved an F1 score of 0.9374 with soft-NMS (threshold 0.7), outperforming the traditional NMS method (F1 score 0.9228).
• The enhanced method demonstrated improved accuracy in segmenting and identifying individual pigs, especially in crowded or overlapping situations.

Conclusions:

• PigMS R-CNN offers a novel and effective instance segmentation approach for adhesive group-housed pig images.
• This research provides a valuable foundation for vision-based, real-time automatic pig monitoring and welfare evaluation systems.