Decoupling Discriminative Attributes for Few-Shot Fine-Grained Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot fine-tuning of vision-language models (VLMs) reduces data needs for downstream tasks.
• VLMs struggle with fine-grained recognition of highly similar subspecies.
• Existing methods fail to capture discriminative features for confusing categories.

Purpose of the Study:

• To develop a hierarchical few-shot fine-tuning framework to address confusion in fine-grained recognition.
• To enhance the ability of VLMs to distinguish between similar subspecies.
• To improve interpretability in few-shot learning.

Main Methods:

• Proposed a two-stage recognition framework: Attribute-Decoupled Discriminator (AttrDD).
• Phase 1: Fine-tuned CLIP to identify Top-K confusing classes.
• Phase 2: Utilized large language models (LLMs) for attribute difference descriptions and attribute-decoupled classification, employing attention adapters for parameter-efficient fine-tuning.

Main Results:

• AttrDD demonstrated superior performance over existing methods on 9 fine-grained recognition benchmarks.
• The framework effectively addresses the severe confusion problem in fine-grained recognition.
• Parameter-efficient fine-tuning was achieved through lightweight attention adapters.

Conclusions:

• The proposed AttrDD framework significantly enhances few-shot fine-tuning for VLMs, particularly for challenging fine-grained recognition tasks.
• Leveraging LLMs for attribute-level distinctions improves discriminative feature extraction.
• AttrDD offers a promising direction for improving VLM performance in specialized recognition scenarios.