| 本文已被：浏览 148次 下载 153次 |
码上扫一扫！ |
|
|
| 样本-特征协同的长尾识别算法 |
|
张恩豪1, 李超华2,3,4, 王志华1, 陈松灿3,4
|
|
1.郑州大学 网络空间安全学院, 河南 郑州 450002;2.西安邮电大学 计算机学院, 陕西 西安 710121;3.南京航空航天大学 计算机科学与技术学院, 江苏 南京 211106;4.模式分析与机器智能工业和信息化部重点实验室(南京航空航天大学), 江苏 南京211106
|
|
| 摘要: |
| 现实场景中的数据分布普遍呈现长尾模式, 导致所训练的深度模型常遭遇头部偏置困境, 即模型偏向于头部类而在尾部类上表现欠佳. 一种简单且有效的策略是通过增广尾部类样本来平衡数据分布. 尽管多数基于该策略的方法能在样本数量上实现重平衡, 但生成的样本仍存在语义漂移和多样性不足的风险, 导致特征分布松散和决策边界有偏, 限制了模型性能. 为此, 提出一种样本-特征协同的长尾识别算法, 旨在构建对数据分布不敏感的分类模型. 具体而言, 在样本层面, 基于傅里叶变换提出“幅值增广”策略, 借助幅值迁移来转换增广样本的风格, 能够在丰富样本多样性的同时保留原有语义信息; 在特征层面, 依据神经坍塌理论提出“特征坍塌”损失, 将类原型对齐至具有最大可分性的等角紧框架, 并促使特征收敛至对应的类原型, 实现最大的类间间隔. 该方法从样本和特征两个层面缓解头部偏置问题, 进而增强类内紧凑性并校准决策边界. 多个基准数据集上的实验结果表明该方法可显著提高长尾识别性能. |
| 关键词: 长尾识别 傅里叶变换 神经坍塌 数据增广 |
| DOI：10.13328/j.cnki.jos.007629 |
| 分类号: |
| 基金项目:国家自然科学基金(62376126); 航空发动机及燃气轮机重大专项基础研究项目(J2019-IV-0018-0086) |
|
| Sample-feature Collaborative Long-tailed Recognition Algorithm |
|
ZHANG En-Hao1, LI Chao-Hua2,3,4, WANG Zhi-Hua1, CHEN Song-Can3,4
|
|
1.School of Cyber Science and Engineering, Zhengzhou University, Zhengzhou 450002, China;2.School of Computer Science & Technology, Xi’an University of Posts & Telecommunications, Xi’an 710121, China;3.College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China;4.MIIT Key Laboratory of Pattern Analysis and Machine Intelligence (Nanjing University of Aeronautics and Astronautics), Nanjing 211106, China
|
| Abstract: |
| Data distributions in real-world scenarios commonly exhibit long-tail patterns, causing deep models to suffer from head bias, where performance is biased toward head classes while tail classes are poorly recognized. A simple and effective strategy to alleviate this problem is to balance data distributions by augmenting tail-class samples. Although most methods following this strategy achieve a quantitative rebalancing, the generated samples often suffer from semantic shift and insufficient diversity, which leads to dispersed feature distributions and biased decision boundaries, thereby limiting overall model performance. To address these issues, this study proposes a sample-feature collaborative learning framework for long-tailed recognition, aiming to construct classification models that are insensitive to data distribution imbalance. At the sample level, a “magnitude augmentation” strategy based on the Fourier transform is introduced, in which amplitude shifting is employed to modify the style of augmented samples, while preserving their original semantic information. At the feature level, a “feature collapse” loss inspired by neural collapse theory is proposed to align class prototypes into an equiangular tight frame with maximum separability and to encourage features to converge toward their corresponding class prototypes, achieving maximum inter-class separation. By jointly addressing head bias from both the sample and feature perspectives, the proposed framework enhances intra-class compactness and calibrates decision boundaries. Experimental results across multiple benchmark datasets demonstrate that the proposed method significantly improves long-tailed recognition performance. |
| Key words: long-tailed recognition (LTR) Fourier transform neural collapse (NC) data augmentation |
