Using Pruning-Based YOLOv3 Deep Learning Algorithm for Accurate Detection of Sheep Face.

Area of Science:

• Computer Vision
• Animal Science
• Machine Learning

Background:

• Accurate sheep identification is crucial for precision animal management and welfare farming.
• Existing methods require optimization for large-scale farm applications.

Purpose of the Study:

• To develop an efficient and accurate sheep face detection method for welfare farming.
• To optimize the YOLOv3 model for improved performance and reduced size.

Main Methods:

• Collected Sunit sheep face images from Inner Mongolia.
• Trained and compared YOLOv3, YOLOv4, Faster R-CNN, and SSD models.
• Optimized YOLOv3 by clustering anchor frames and model compression, achieving YOLOv3-P.

Main Results:

• Optimized YOLOv3 (YOLOv3-P) achieved a mean average precision (mAP) of 97.2%.
• Model size was reduced to 1/4 of the original YOLOv3.
• 10-fold cross-validation yielded an mAP of 96.84%.

Conclusions:

• Clustering anchor boxes and compressing the YOLOv3 model is effective for sheep identification.
• YOLOv3-P offers low memory requirements, high accuracy, and fast recognition.
• The method has significant applications in precision animal management and welfare farming.