# Anisotropic Diffusion Probabilistic Model for Imbalanced Image Classification

**Authors:** Jingyu Kong, Yuan Guo, Yu Wang, Yuping Duan  
**arXiv:** [2409.14313](https://arxiv.org/abs/2409.14313)  
**Date:** 2024-09-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2409.14313v1/

## Key points

- The paper addresses the challenge of long-tailed data distribution in real-world datasets, which limits generalization ability in machine learning models.
- The authors introduce the Anisotropic Diffusion Probabilistic Model (ADPM) to improve classification accuracy for tail classes in imbalanced image classification problems.
- ADPM controls diffusion speed based on class sample distribution, enhancing denoiser's classification accuracy and addressing traditional model struggles with rare class classification.
- A theoretical framework is proposed for selecting noise levels tailored to each category, effectively tackling imbalanced classification challenges and improving model performance.
- Global and local image priors are integrated into the forward process to enhance spatial discriminative abilities, while semantic-level contextual information is incorporated into the reverse process to boost overall model discriminative power and robustness.
- Extensive comparisons with state-of-the-art methods on medical benchmark datasets demonstrate significant improvements in F1-scores for rare classes using ADPM.

## Summary

The paper "Anisotropic Diffusion Probabilistic Model for Imbalanced Image Classification" by Jingyu Kong, Yuan Guo, Yu Wang, and Yuping Duan addresses the challenge of long-tailed data distribution in real-world datasets. This often leads to limited generalization ability in machine learning models. The authors introduce the Anisotropic Diffusion Probabilistic Model (ADPM) as a solution to improve classification accuracy for tail classes in imbalanced image classification problems. The ADPM is built upon the foundation of Denoising Diffusion Probabilistic Models (DDPM), which leverage stochastic differential equation theory and have shown promising results in image classification tasks. However, traditional diffusion probabilistic models struggle with accurately classifying rare classes within imbalanced datasets. To address this issue, the ADPM incorporates a novel approach that controls the diffusion speed of different class samples based on their distribution during the forward process. This strategy enhances the denoiser's classification accuracy in the reverse process. Moreover, the authors propose a theoretical framework for selecting noise levels tailored to each category in the diffusion process. By leveraging error analysis theory, they effectively tackle imbalanced classification challenges and improve model performance. Additionally, global and local image priors are integrated into the forward process to enhance spatial discriminative abilities while semantic-level contextual information is incorporated into the reverse process to boost overall model discriminative power and robustness. Through extensive comparisons with state-of-the-art methods on four medical benchmark datasets – including skin lesion datasets PAD-UFES and HAM10000 – the authors demonstrate the effectiveness of ADPM in handling long-tail data distributions. Results show significant improvements in F1-scores for rare classes compared to traditional diffusion probabilistic models, with an increase of 4% and 3% on PAD-UFES and HAM10000 datasets respectively. Overall, this work highlights how anisotropic diffusion modeling can significantly enhance classification accuracy for challenging imbalanced image datasets while maintaining high accuracy for dominant classes.