基于分层动态掩码视觉Transformer的长尾图像识别研究OA
Hierarchical Dynamic Masked Vision Transformer for Long-Tail Image Recognition
现实世界中的许多分类任务天然具有层次化语义结构,当这些层次化分类任务同时面临长尾分布时,识别难度进一步加剧.现有的长尾学习方法大多将类别视为扁平结构,未能充分利用层次标签中固有的语义关联.传统的层次分类方法通常假设类别分布均衡,对长尾分布缺乏鲁棒性.为解决这一问题,本文提出了一种基于分层动态掩码视觉 Transformer 的长尾图像识别算法(hierarchical dynamic masked vision transformer,HDMViT),针对标签分布极不均衡问题,对数据集的父类层标签进行了重新构建以优化层次结构的表达能力.同时,在视觉Transformer 模型中引入层次化双重分类头结构,两个分类头均基于 Transformer 编码器输出的同一特征向量进行决策:粗粒度分类头(父类)通过全连接层直接建模父类分布,细粒度分类头(子类)则进一步设计动态 Logit 掩码机制来过滤无关特征干扰,确保模型能够更专注地学习目标类别的关键特征,从而更有效地利用细粒度和粗粒度的语义信息.实验结果表明,该方法在多个长尾数据集上均表现优异.
Many classification tasks in the real world inherently possess hierarchical semantic structures.When these hierarchical classification tasks simultaneously face long-tailed distributions,the difficulty of recognition is further exacerbated.Existing long-tailed learning methods mostly treat categories as flat structures and fail to fully utilize the inherent semantic correlations in hierarchical labels.Traditional hierarchical classification methods,on the other hand,usually assume balanced category distributions and lack robustness to long-tailed distributions.To address this problem,this paper proposes a long-tailed image recognition algorithm based on a Hierarchical Dynamic Masked Vision Transformer(HDMViT).Aiming at the extremely imbalanced label distribution,the parent-level labels of the dataset are reconstructed to enhance the expressive capability of the hierarchical structure.Proposed methodology reconstructs parent-level labels to strengthen hierarchical representation,specifically targeting extreme label imbalance.Meanwhile,a hierarchical dual-classification-head structure is introduced into the Vision Transformer model,in which both classification heads make decisions based on the same feature vector output by the Transformer encoder:the coarse-grained(parent-level)head directly models the parent-class distribution through fully connected layers,while the fine-grained(child-level)head further incorporates a dynamic logit masking mechanism to filter out irrelevant feature interference,ensuring that the model learns the key features of the target classes in a more focused manner,thereby making more effective use of both fine-grained and coarse-grained semantic information.Experimental results show that the proposed method achieves excellent performance on multiple long-tailed datasets.
汪世豪;毛煜;林耀进
闽南师范大学计算机学院,福建 漳州 363000||闽南师范大学数据科学与智能应用福建省高校重点实验室,福建 漳州 363000闽南师范大学计算机学院,福建 漳州 363000||闽南师范大学数据科学与智能应用福建省高校重点实验室,福建 漳州 363000闽南师范大学计算机学院,福建 漳州 363000||闽南师范大学数据科学与智能应用福建省高校重点实验室,福建 漳州 363000
信息技术与安全科学
深度学习长尾学习不平衡数据视觉Transformer层次分类
deep learninglong-tail learningimbalanced datavision transformerhierarchical classification
《南京师范大学学报(工程技术版)》 2026 (3)
22-30,9
国家自然科学基金资助项目(62576158)、福建省自然科学基金重点项目(2026J002035)、福建省自然科学基金面上项目(2026J001987).
评论