LAD-RCNN: A Powerful Tool for Livestock Face Detection and Normalization.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Animal Science

Background:

• Standardized large-scale livestock farming necessitates advanced AI for animal monitoring.
• Existing animal face identification systems lack specialized face normalization, hindering performance.
• Human face normalization techniques are unsuitable for livestock due to uncooperative subjects and arbitrary image angles.

Purpose of the Study:

• To develop a robust livestock face normalization method capable of handling arbitrary face directions.
• To improve the accuracy and efficiency of livestock face detection and identification systems.
• To address the gap in specialized livestock face normalization research.

Main Methods:

• Development of a lightweight angle detection and region-based convolutional network (LAD-RCNN).
• Implementation of a novel rotation angle coding method for simultaneous angle and location detection.
• Integration of image enhancement techniques to boost LAD-RCNN performance.
• Evaluation on diverse datasets, including goat and infrared imagery.

Main Results:

• LAD-RCNN achieved over 97% average precision in face detection across test datasets.
• Detected rotation angle deviations were less than 6.42° compared to ground truth.
• The model processes images rapidly, taking only 13.7 ms per image on a single RTX 2080Ti GPU.

Conclusions:

• LAD-RCNN demonstrates excellent performance in livestock face recognition and direction detection.
• The developed method is highly suitable for livestock face detection and normalization tasks.
• This research provides a foundational solution for improving AI applications in livestock management.