Abstract: Complex intra-class feature distributions in fine-grained image classification pose significant challenges to discriminative feature learning. Existing proxy-based metric learning methods typically employ a fixed number of proxies for all classes, making it difficult to accommodate differences in intra-class structural complexity and limiting their ability to model complex intra-class distributions. To address this issue, this paper proposes an Adaptive Multi-Proxy Metric Learning (AMPML) method. The method uses historical sample-to-proxy assignment probabilities to update proxy usage rates in a smoothed manner, reflecting how frequently each proxy is used by samples of its class. It then dynamically selects effective proxies for each class according to an activation threshold, allowing the number and set of proxies involved in optimization to adapt to changes in the class feature distribution and differences in intra-class structural complexity. On this basis, potential field theory is used to model attraction between samples and proxies of the same class and repulsion between samples and proxies of different classes, thereby optimizing the spatial distribution of samples and proxies. Potential-field interactions are calculated only between samples and currently effective proxies; inactive proxies do not contribute to the potential energy. Consequently, adjusting the proxy set at the class level directly changes the range of potential-field interactions and the resulting optimization constraints for samples of each class. To address the limited ability of sample-to-proxy constraints to model local boundary relationships, a hard-sample constraint is further introduced to improve the intra-class compactness of boundary samples. Specifically, hard samples are selected according to the model’s predicted confidence for their ground-truth class, and a distance constraint is constructed between hard samples from the same class to supplement the modeling of local intra-class relationships by the sample-to-proxy potential field. Experiments are conducted on three public fine-grained datasets: CUB-200-2011, Stanford Dogs, and FGVC-Aircraft. Classification results obtained with different backbone networks are compared, and ablation experiments are used to analyze the effects of individual components on classification accuracy and their combined contribution. The results show that the proposed method effectively optimizes the feature-space distribution, improves the modeling of complex intra-class structures, and achieves better classification performance than several existing metric learning methods. Ablation experiments and visualization analyses further verify the effectiveness of the components and their complementary roles.

摘要： 细粒度图像分类中复杂的类内特征分布给判别性特征学习带来了较大挑战。现有基于代理的度量学习方法通常采用固定数量代理表示所有类别，难以适应不同类别内部结构复杂度的差异，从而限制了对复杂类内分布的刻画能力。为此，提出一种自适应多代理度量学习方法（Adaptive Multi-Proxy Metric Learning，AMPML）。该方法根据样本与代理的历史匹配概率平滑更新代理使用率，以反映各代理在类别样本中的使用情况；随后依据激活阈值动态筛选各类别的有效代理，使参与优化的代理数量与集合随类别特征分布变化而调整，从而适应不同类别的类内结构复杂度。在此基础上，根据势场理论建模同类样本与代理之间的吸引关系以及异类样本与代理之间的排斥关系，以优化代理点与样本的空间分布。势场计算仅涉及样本与当前有效代理之间的相互作用，未激活代理不参与势能计算。由此，类别级代理规模调整直接改变势场作用范围，使不同类别的样本接受与其有效代理集合对应的优化约束。同时，针对样本—代理约束对局部边界刻画不足的问题，采用困难样本约束增强边界样本的类内紧凑性。具体而言，根据模型对真实类别的预测置信度筛选困难样本，在同类别困难样本之间构建距离约束，以补充样本—代理势场对局部类内关系的建模。最后，在CUB-200-2011、Stanford Dogs和FGVC-Aircraft三个公开细粒度数据集上进行实验验证。同时比较不同主干网络的分类结果，并通过消融实验分析各模块对分类准确率的影响及其协同关系。实验结果表明，所提出方法能够有效优化特征空间分布，提高复杂类内结构的建模能力，并取得优于多种现有度量学习方法的分类性能。消融实验和可视化分析进一步验证了各组成模块的有效性及其协同作用。自适应多代理度量学习的细粒度图像分类