Feature relocation network for fine-grained image classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Fine-grained image classification faces challenges due to subtle inter-class similarities and large intra-class variances.
• Existing methods often rely on manual annotations or attention mechanisms, which struggle to effectively separate discriminative from distractive features.

Purpose of the Study:

• To develop a novel network, the Feature Relocation Network (FRe-Net), for improved fine-grained image classification.
• To effectively learn local discriminative features and remove distractive features by leveraging multi-stage feature characteristics.

Main Methods:

• Proposing a FRe-Net comprising a distractive feature learning module and a relocated high-level feature learning module.
• Introducing a distractive loss (L_distractive) that utilizes the differences between low-level and high-level features to guide attention.
• Enhancing middle-level features via an attention module and subtracting learned distractive features to refine local discriminative features.

Main Results:

• FRe-Net demonstrated state-of-the-art performance on three widely-used benchmark datasets for fine-grained image classification.
• The joint optimization of the distractive and relocated feature learning modules proved mutually beneficial.

Conclusions:

• The proposed FRe-Net effectively addresses the limitations of existing methods in fine-grained image classification.
• The feature relocation strategy significantly enhances the model's ability to learn discriminative features, validating its effectiveness.