Multiresolution Discriminative Mixup Network for Fine-Grained Visual Categorization.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Fine-grained visual categorization (FGVC) is difficult due to subtle inter-class differences in local regions.
• Existing methods use high-resolution images or regularization techniques like mixup, but face challenges like high computational costs and manifold intrusion.

Purpose of the Study:

• To develop an efficient and accurate method for FGVC that overcomes the limitations of current approaches.
• To improve the learning of local detail features and reduce computational expenses.

Main Methods:

• Proposing a Multiresolution Discriminative Mixup Network (MRDMN).
• Implementing a discriminative mixup strategy that mixes discriminative image regions instead of entire images to prevent manifold intrusion.
• Designing a resolution-based distillation strategy to transfer multiresolution features to a low-resolution network.

Main Results:

• The MRDMN effectively learns local detail features, leading to more precise categorization.
• The resolution-based distillation speeds up testing and boosts categorization accuracy.
• Experiments show MRDMN outperforms competitive approaches on multiple benchmark datasets (CUB-200-2011, Stanford-Cars, Stanford-Dogs, Food-101, iNaturalist 2017) with reduced computation time.

Conclusions:

• MRDMN offers a superior solution for FGVC by addressing manifold intrusion and computational cost issues.
• The proposed discriminative mixup and resolution-based distillation strategies are effective for enhancing both accuracy and efficiency in visual categorization.