Multi-Granularity Part Sampling Attention for Fine-Grained Visual Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Fine-grained visual classification (FGVC) faces challenges due to intra-class variation and inter-class similarity.
• Existing methods often struggle to capture rich shape information by relying on rectangular bounding boxes or standard attention mechanisms.
• There is a need for methods that can extract discriminative part features more effectively.

Purpose of the Study:

• To propose a novel network, the Multi-Granularity Part Sampling Attention (MPSA) network, for addressing FGVC challenges.
• To enhance the extraction of semantic part features, focusing on shape and scale variations.
• To improve the overall performance and robustness of fine-grained visual classification models.

Main Methods:

• Introduced a multi-granularity part retrospect block to extract part information at different scales.
• Developed a part sampling attention mechanism for comprehensive sampling of implicit semantic parts with varied shapes.
• Implemented a part dropout strategy to mitigate overfitting.
• Proposed a multi-granularity fusion method utilizing gradient class activation maps to emphasize foreground features and reduce background noise.

Main Results:

• The MPSA network achieved state-of-the-art performance on four standard fine-grained visual classification benchmarks.
• The proposed methods effectively captured detailed shape information and discriminative part features.
• The multi-granularity approach enhanced feature representation and classification accuracy.

Conclusions:

• The MPSA network offers a significant advancement in fine-grained visual classification.
• The novel attention and fusion mechanisms provide a more robust and accurate approach to identifying subtle visual differences.
• The method's effectiveness is validated by its superior performance on established datasets.