Webly Supervised Fine-Grained Classification by Integrally Tackling Noises and Subtle Differences.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Webly-supervised fine-grained visual classification (WSL-FGVC) faces challenges with noisy web image labels and distinguishing subtle inter-class variations.
• Current WSL-FGVC methods primarily address image-level noise, neglecting pixel-level feature mining crucial for fine-grained distinctions.

Purpose of the Study:

• To develop an integrated framework for WSL-FGVC that simultaneously handles label noise and extracts subtle pixel-level discriminative cues.
• To improve the robustness and accuracy of fine-grained visual classification using web-scraped data.

Main Methods:

• Proposed a bag-level top-down attention framework to process groups of images from the same class.
• Extracted high-level semantic information from image bags to guide the mining of discriminative regions at various scales within individual images.
• Introduced attention-based mechanisms for robust bag-level fusion and an attention loss to refine attention map learning.

Main Results:

• The proposed framework effectively tackles label noise and mines subtle visual cues.
• Demonstrated superior performance compared to state-of-the-art methods across four WSL-FGVC benchmark datasets (Web-Aircraft, Web-Bird, Web-Car, WebiNat-5089).

Conclusions:

• The bag-level top-down attention framework offers an effective and integrated solution for WSL-FGVC.
• The method's ability to simultaneously address noise and mine fine-grained details significantly advances the field.