基于ViT的细粒度图像分类

Fine-grained visual classification based on vision transformer

为解决细粒度图像分类任务存在类内差异性和类间相似性大的问题,提出一种基于Vision Transformer(ViT)的细粒度图像分类方法。采取ViT作为特征编码网络,获取图像的全局特征表示;设计多级区域选择模块,捕捉细微的具有可判别性的层级化信息;利用一个简单且有效的中心损失函数,缩短深层特征与相应类中心在特征空间中的距离。在图像级标签的监督下,实现端到端的训练。结果在CUB-200-2011、NABirds以及Stanford Cars数据集上分别达到90.1%、90.2%和93.7%的分类准确率,超越当前最优算法。

To address the problem of subtle intra-class similarities and big inter-class variances fronted by in the fine-grained visual classification(FGVC)task,a method based on Vision Transformer(ViT)architecture was proposed for FGVC.Speci-fically,ViT as features encoding network was used to extract image representation.Multi-regions selection module(MRSM)was designed to capture fine-grained discriminative and hierarchical information.The center loss was introduced to close the distance between depth features and class centers.Under the supervision of image-level labels,the entire network was optimized in an end-to-end manner.The results achieve 90.1%,90.2%and 93.7%classification accuracies on the CUB-200-2011,NABirds and Stanford Cars datasets,respectively,surpassing the current best algorithm.