An Improved Lightweight Model for Protected Wildlife Detection in Camera Trap Images.

Area of Science:

• Ecology
• Computer Science
• Artificial Intelligence

Background:

• Effective wildlife monitoring is vital for biodiversity conservation.
• Current deep learning models struggle with detecting rare species and have high computational demands, limiting edge device deployment.
• There is a need for efficient and accurate wildlife detection models for ecological observation.

Purpose of the Study:

• To propose YOLO11-APS, an improved lightweight deep learning model for protected wildlife detection.
• To enhance feature extraction and reduce computational costs for deployment on edge devices.
• To achieve a balance between detection accuracy and model complexity.

Main Methods:

• Integration of the self-Attention and Convolution (ACmix) module, Partial Convolution (PConv) module, and SlimNeck paradigm into the YOLO11n architecture.
• Development of a lightweight model for protected wildlife detection.
• Experimental evaluation of detection performance and model complexity.

Main Results:

• YOLO11-APS achieved superior detection performance: 92.7% precision, 87.0% recall, 92.6% mAP@0.5, and 62.2% mAP@0.5:0.95.
• Model lightweighting resulted in a 10.1% reduction in parameters, 11.1% in FLOPs, and 9.5% in model size.
• YOLO11-APS outperformed existing lightweight detection models in accuracy and complexity.

Conclusions:

• YOLO11-APS offers an optimal balance between accuracy and model complexity for wildlife detection.
• The model demonstrates strong transferability and robustness on unseen wildlife data.
• This work provides an efficient deep learning tool for automated wildlife monitoring and intelligent ecological sensing systems.