A Lightweight Automatic Wildlife Recognition Model Design Method Mitigating Shortcut Learning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Ecology

Background:

• Wildlife recognition from camera trap images is hindered by complex environments and similar backgrounds, leading to shortcut learning in deep learning models.
• Existing models often exhibit poor generality and performance due to over-reliance on background features.

Purpose of the Study:

• To develop a data augmentation strategy to improve the background richness and suppress background information in camera trap images.
• To create a lightweight deep learning model for real-time wildlife recognition on edge devices.

Main Methods:

• Proposed a data augmentation strategy combining image synthesis (IS) and regional background suppression (RBS).
• Developed a model compression strategy using adaptive pruning (genetic algorithm-based pruning and adaptive batch normalization - GA-ABN) and knowledge distillation (mean square error - MSE loss).

Main Results:

• The data augmentation strategy effectively enriches background scenes and suppresses existing background information, guiding models to focus on wildlife.
• The compressed lightweight model achieved a reduction in computational effort for wildlife recognition with only a 4.73% loss in accuracy.
• Extensive experiments validated the advantages of the proposed methods for real-time wildlife monitoring.

Conclusions:

• The integrated data augmentation and model compression strategies significantly improve the generality and performance of wildlife recognition models.
• The developed lightweight model is suitable for real-time wildlife monitoring applications on edge devices, balancing accuracy and computational efficiency.