基于多尺度判别模块的细粒度图像分类

Fine-grained Image Classification Based on Multi-scale Discrimination Module

为了提高混合架构在细粒度图像分类领域的性能,设计提出了一个判别模块(DM),该模块包括两个部分:判别性特征选择(DFS)和多尺度特征聚合(MFA)。DFS模块通过选取Vision Transformer(ViT)中不同注意力头中Top-K个鸟类物种的代表性特征以关注不同区域特征,促进不同判别区域的协同效应,同时减少特征冗余。MFA模块聚合不同尺度的鸟类判别性特征信息。在开源的鸟类细粒度数据集上进行了实验证明,并与现有方法进行了比较。实验结果表明,所提出的模块在鸟类细粒度图像识别方面取得了一定的改进。

In order to improve the performance of hybrid architecture in the field of fine-grained image classification,a Discriminant Module(DM)is proposed,which consists of two parts:Discriminant Feature Selection(DFS)and Multi-scale Feature Aggregation(MFA).By selecting the representative features of Top-K bird species in different attentional heads in Vision Transformer(ViT),the DFS module pays attention to the characteristics of different regions,promotes the synergistic effect of different discriminating regions,and reduces the feature redundancy.The MFA module aggregates discriminant characteristics information of birds at different scales.The experimental proof is carried out on an open source bird fine-grained dataset and compared with existing methods.The experimental results show that the proposed module has achieved some improvement in bird fine-grained image recognition.