An Image Detection Model for Aggressive Behavior of Group Sheep

Area of Science:

• Agricultural Science
• Computer Vision
• Animal Behavior

Background:

• Sheep aggression detection is vital for welfare in large-scale farming.
• Current methods often lack lightweight models suitable for group detection.
• Existing video detection models struggle with accurate localization of aggressive individuals.

Purpose of the Study:

• To propose a lightweight network model for detecting aggression in group sheep.
• To enhance the GhostNet architecture for improved feature extraction in sheep behavior analysis.
• To achieve accurate and real-time detection of aggressive sheep within a group.

Main Methods:

• Utilized GhostNet as the core feature extraction network.
• Incorporated PWConv and Channel Shuffle operations into GhostConv modules for enhanced information exchange.
• Introduced the Inverted-GhostBottleneck module with an inverted residual structure to increase feature map information.

Main Results:

• The improved GhostNet achieved 94.7% Precision and 90.7% Recall.
• The model size was reduced to 62.7% compared to YOLOv5.
• Successfully addressed limitations of previous video detection models in accurate localization.

Conclusions:

• The enhanced GhostNet model offers a highly effective and lightweight solution for group sheep aggression detection.
• The proposed model demonstrates superior performance over the original GhostNet and addresses key limitations in current methods.
• Enables real-time, accurate identification of aggressive behaviors, contributing to improved animal welfare.