Developing and Evaluating Poultry Preening Behavior Detectors via Mask Region-Based Convolutional Neural Network.

Area of Science:

• Agricultural Science
• Computer Vision
• Animal Behavior

Background:

• Automated monitoring of poultry behavior is crucial for welfare and productivity.
• Existing tools lack precision for automated preening behavior detection in chickens.

Purpose of the Study:

• To develop and evaluate Mask R-CNN based detectors for automated poultry preening behavior identification.
• To assess the performance of Mask R-CNN in detecting individual hen preening in group settings.

Main Methods:

• Utilized Mask R-CNN for developing preening behavior detectors from surveillance images of brown hens.
• Evaluated detector performance using metrics like MIOU, precision, recall, F1 score, and processing speed.
• Investigated the impact of network architecture (ResNets), transfer learning, and image preprocessing on detection.

Main Results:

• Mask R-CNN achieved high performance with 94.2% accuracy and 86.5% F1 score.
• Processing speed was 380.1 ms per image, suitable for real-time monitoring.
• Transfer learning and image resizing strategies were explored with varying impacts on performance.

Conclusions:

• The Mask R-CNN preening behavior detector is a viable tool for automated, precise monitoring of individual hen preening.
• This technology can significantly aid in understanding and managing poultry welfare and behavior in group environments.