Robust Fine-Grained Visual Recognition With Neighbor-Attention Label Correction.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning for fine-grained visual recognition requires extensive, accurate annotations.
• Real-world data collection often introduces label noise, hindering model performance.
• Existing methods struggle with the inherent noise in fine-grained datasets.

Purpose of the Study:

• To address the challenge of label noise in deep model training for fine-grained visual recognition.
• To propose a novel method for correcting noisy labels during the training phase.
• To enhance the robustness and accuracy of deep learning models in the presence of label noise.

Main Methods:

• Proposed the Neighbor-Attention Label Correction (NALC) model for automated label correction.
• Utilized a meta-learning framework with a validation batch to correct training batch labels.
• Introduced a nested optimization algorithm to improve meta-learning efficiency.
• Implemented NALC to refine label accuracy within training batches.

Main Results:

• Significantly improved label accuracy from 70% to over 98%.
• Outperformed existing methods by up to 13.4% in mean Average Precision (mAP) on fine-grained image retrieval tasks.
• Achieved a 7.8% improvement in mean Intersection over Union (mIOU) on noisy semantic segmentation datasets.
• Demonstrated robustness against various simulated and practical noise types.

Conclusions:

• NALC effectively corrects noisy labels, enhancing learned image representations.
• The proposed method offers substantial performance gains across diverse fine-grained visual recognition tasks.
• NALC provides a robust solution for training deep learning models with noisy datasets.