Dual-Dependency Attention Transformer for Fine-Grained Visual Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Visual Transformers (ViTs) are prevalent in visual tasks like fine-grained visual classification (FGVC).
• The core self-attention mechanism in ViTs incurs quadratic computational and memory costs.
• Existing sparse-attention and local-attention methods are inadequate for FGVC's need for dense features and global dependency modeling.

Purpose of the Study:

• To develop an efficient attention mechanism for Visual Transformers suitable for fine-grained visual classification.
• To address the computational and memory complexity limitations of standard self-attention in ViTs.
• To enhance the modeling of discriminative cues for improved FGVC performance.

Main Methods:

• Proposing a dual-dependency attention transformer that decouples global token interactions.
• Implementing a position-dependency attention pathway using grouped attention.
• Utilizing a semantic dependency attention pathway with dynamic central aggregation.
• Developing discriminative enhancement strategies for sensitive cue tracking using knowledge-based representations.

Main Results:

• The dual-dependency attention transformer achieves linear computational complexity.
• The model demonstrates suitability for fine-grained image classification tasks.
• Experimental results on NABIRDS, CUB, and DOGS datasets validate the proposed approach.

Conclusions:

• The proposed dual-dependency attention transformer offers an effective solution for fine-grained visual classification.
• This method balances computational efficiency with high performance in FGVC.
• The approach enhances semantic modeling and discriminative cue tracking, outperforming existing methods for FGVC.