Mixed Mutual Transfer for Long-Tailed Image Classification.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Real-world datasets frequently exhibit long-tailed distributions, characterized by numerous majority (head) classes and scarce minority (tail) classes.
• This data imbalance creates disparities, negatively impacting deep learning network performance.
• Existing knowledge transfer methods often lack diversity, focusing solely on augmenting minority classes.

Purpose of the Study:

• To introduce Mixed Mutual Transfer (MMT), a novel method for long-tailed classification.
• To facilitate mutual knowledge transfer between head and tail classes by blending samples.
• To enhance the performance of deep networks on imbalanced datasets.

Main Methods:

• Developed Mixed Mutual Transfer (MMT) for long-tailed classification.
• Employs a uniform sampler for head samples and a differential sampler for tail samples.
• Generates new, diverse samples by blending head and tail class data for network training.

Main Results:

• MMT diversifies both tail and head classes, addressing information disparity.
• The method significantly enhances the performance of both majority and minority classes.
• Experimental results demonstrate superior performance compared to existing methods on imbalanced datasets.

Conclusions:

• Mixed Mutual Transfer (MMT) is an effective strategy for improving long-tailed deep network performance.
• The proposed blending technique successfully mitigates data imbalance issues.
• MMT offers a promising approach for diverse and robust deep learning models.