Augmentation variational feature embedding for sample specificity learning
Yue Xia Hu, Jianwei Yang · Journal of Electronic Imaging · 2025
Sample specificity learning aims to consider each individual sample as an independent class and explore the underlying visual similarity relationships among samples, thereby learning discriminative feature embeddings without relying on category labels. However, existing data augmentation strategies not only provide limited supervision signals, which lead to model overfitting, but also result in the learning of poor discriminative features. In this article, we propose an augmentation variation feature embedding method for sample specificity learning. Specifically, instance feature intra-variation utilizes residual variation within a single instance feature to generate augmented features in geometric space, providing positive supervision while avoiding the introduction of noise. Instance feature inter-variation generates a mixed feature between two instance features in geometric space, encouraging the model to learn potential instance-to-instance similarity relationships. Extensive experiments on three fine-grained image datasets demonstrate the superior discriminative power and generalization ability of our proposed method compared with existing approaches.