.png){kind=icon}
返回
DDFA: a displacement and diffusion-based feature augmentation method for imbalanced image recognition
.png){kind=icon}
PRE
AI
DOI：10.1007/s00371-024-03673-z.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
Data-driven computer vision methods have achieved great success in multiple fields, but how to learn a balanced classifier on imbalanced distribution remains a great challenge for the data-driven method. The key issue behind long-tailed distribution is the information insufficiency of tail classes. Recent works adopted information augmentation (IA) to generate new tail class samples for mitigating this issue. However, the existing IA methods usually ignore feature drift, which hurt the decision boundary. Additionally, these methods only leverage limited information to generate samples, which cannot guarantee the quality of sample generation. To address these issues, we propose a displacement and diffusion-based feature augmentation (DDFA) method for learning a balanced model on imbalanced training distribution. Firstly, the feature reverse displacement module performs feature displacement on original tail features. It can mitigate the feature drift between the head class and the tail class. Subsequently, a long-tailed diffusion model is proposed to generate high-quality tail class samples with diversity and fidelity, which can mitigate the information insufficiency issue of tail class. Finally, the original samples and generated samples are combined in the feature space to promote balanced classifier learning. Experimental results on four challenging datasets demonstrate the effectiveness of the proposed DDFA method. The code is available at: https://github.com/wzh-why/DDFA.
Keyword:
Feature augmentation
Diffusion model
Long-tailed learning
Computer vision
期刊
IF:
2.9
论文数:
4.6K
被引数:
6.5K
机构
引用论文
CoRe: Contrastive and Restorative Self-Supervised Learning for Surface Defect InspectionCoRe: 用于表面缺陷检测的对比和恢复性自监督学习
PREAI
分享收藏Generation of hydroxyl radicals by urban suspended particulate air matter. The role of iron ions城市悬浮颗粒物产生羟基自由基。铁离子的作用:
PREAI
分享收藏A deep learning system for detecting diabetic retinopathy across the disease spectrum用于检测整个疾病谱中糖尿病视网膜病变的深度学习系统.png)
NATURE COMMUNICATIONS
OAAI
NATURE COMMUNICATIONSIF15.7
OAAI分享收藏分享收藏Prevalence and determinants of malaria among children in Zambézia Province, Mozambique赞比西亚省莫桑比克儿童中疟疾的患病率及其影响因素OAAI
OAAI分享收藏
查看更多内容