Cross-Part Learning for Fine-Grained Image Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Fine-grained visual classification (FGVC) relies on identifying subtle, distinctive features.
• Existing methods often combine features from different image parts without considering their interactions.
• The contribution of significant parts to sub-category prediction needs to be prioritized.

Purpose of the Study:

• To propose a novel weakly supervised method for exploring cross-learning among multi-regional features in FGVC.
• To enhance category prediction by considering interactions between different image parts.
• To improve the contribution of significant parts to sub-category decisions.

Main Methods:

• Introduction of a Cross-Part Convolutional Neural Network (CP-CNN).
• Implementation of a context transformer for joint feature learning across different parts, guided by a 'navigator' part.
• Design of a part proposal generator (PPG) with feature enhancement blocks to precisely locate subtle discriminative parts and alleviate scale variations.

Main Results:

• The proposed CP-CNN method effectively explores cross-learning among multi-regional features.
• The navigator mechanism allows high-confidence parts to guide lower-confidence parts, retaining complementary information.
• Experiments on three benchmark datasets show consistent outperformance over state-of-the-art methods.

Conclusions:

• Considering interactions between image parts significantly improves FGVC performance.
• The CP-CNN, guided by a navigator and utilizing a PPG, offers a robust approach for weakly supervised FGVC.
• The method effectively handles scale variations and improves the accuracy of fine-grained classification.