KT-YOLO: A multi-convolution kernel collaboration model for dense Hu sheep behavior detection.

Area of Science:

• Computer Vision
• Deep Learning
• Animal Behavior Analysis

Background:

• Dense sheep distribution and imbalanced datasets challenge accurate behavior detection in intensive farming.
• Existing models often require over-parameterization, increasing computational load and limiting practical use.
• High misclassification rates hinder effective monitoring and management of sheep in farming environments.

Purpose of the Study:

• To develop an accurate and efficient deep learning model for detecting Hu sheep behaviors in intensive farming.
• To address challenges posed by dense sheep populations, occlusion, and imbalanced behavioral data.
• To introduce a novel dataset and model architecture for improved sheep behavior analysis.

Main Methods:

• Introduction of the Hu Sheep Behavior Dataset (HSBD) with 280 images of 6,766 Hu sheep exhibiting four key behaviors.
• Development of the KT-YOLO model, enhancing YOLOv8n with Kernel-Team Fusion (KTF) using varied convolution kernel sizes.
• Implementation of SlideLoss function to counteract accuracy degradation from imbalanced behavioral categories.

Main Results:

• KT-YOLO achieved a mean Average Precision (mAP50) of 86.4%, a 6.3% improvement over YOLOv8n.
• SlideLoss further improved performance by an additional 1%.
• KT-YOLO demonstrated superior performance in dense Hu sheep behavior detection compared to YOLOv13n.

Conclusions:

• The developed Hu Sheep Behavior Dataset (HSBD) and KT-YOLO model significantly enhance the accuracy and efficiency of dense Hu sheep behavior detection.
• The Kernel-Team Fusion (KTF) method effectively addresses multi-scale features and occlusion challenges in sheep detection.
• This research highlights the practical value and potential of deep learning in optimizing intensive farming management.