Abstract: Fine-grained visual classification often leads to unstable recognition results due to the significant visual differences among objects within the same sub-class. To address this issue, this paper proposes a novel contrastive fine-grained visual classification network that suppresses intraclass variance by capturing the semantic consistency of visual changes within the same class. Specifically, it first uses spatial attention maps to embed discriminative regions to distinguish different sub-classes, and in this process, combines a feature selection self-attention module to enhance the model's focus on key features and reduce background interference, thereby effectively enhancing feature representation and extracting detailed information. Experimental results show that the proposed model has achieved significant performance improvements in multiple vehicle classification tasks, verifying the effectiveness and universality of the proposed method.

Key words: attention mechanism; feature selection; semantic enhancement; vehicle classification

摘要： 细粒度视觉分类在同一子类对象在视觉外观上差异显著，容易导致识别结果不稳定。 为缓解这一问题，文章提出了一种新颖的对比细粒度视觉分类网络，通过捕捉类内图像视觉 变化的语义一致性来抑制类内方差。具体而言，首先利用空间注意力图嵌入判别性区域以区 分不同子类，并在此过程中结合特征选择注意力模块，提升模型对关键特征的关注度，减少 背景干扰，从而有效增强特征表达并提取细节信息。实验结果表明该研究的模型在多个分类 任务中均取得了显著性能提升，验证了所提方法的有效性与普适性。

关键词: 注意力机制；特征选择；语义增强；车辆分类