A Fine-Grained Bird Classification Method Based on Attention and Decoupled Knowledge Distillation.

Area of Science:

• Computer Vision
• Ecological Monitoring
• Machine Learning

Background:

• Accurate bird classification is crucial for ecological monitoring.
• Bird image classification faces challenges like high intraclass variance, low inter-class variance, and low model efficiency.
• Existing methods struggle with fine-grained distinctions and computational demands.

Purpose of the Study:

• To develop a fine-grained bird classification method that addresses challenges in accuracy and efficiency.
• To improve bird recognition for ecological applications through advanced deep learning techniques.
• To achieve efficient model compression without sacrificing classification performance.

Main Methods:

• An attention-guided data augmentation method was proposed to focus on key bird part regions, enhancing feature learning.
• A localization-recognition approach was integrated to predict bird categories using finer features, reducing background noise interference.
• Decoupled knowledge distillation was employed for efficient model compression, separately distilling target and non-target class knowledge.

Main Results:

• The proposed method achieved an 87.6% success rate in bird classification.
• The model demonstrated significant efficiency, with 67% fewer parameters and only 1.2 G of computation.
• Model inference speed was substantially improved while maintaining high accuracy.

Conclusions:

• The novel method effectively improves fine-grained bird classification accuracy and efficiency.
• Attention mechanisms and decoupled knowledge distillation offer a promising approach for ecological monitoring and computer vision tasks.
• The developed model provides a computationally efficient solution for bird recognition with high performance.