Context-aware contrastive learning via structural harmony preservation for generalized category discovery

This paper addresses the challenge of Generalized Category Discovery (GCD), where models must recognize both known and unknown categories in an open-world setting. We propose a novel Context-Aware Contrastive Learning (CACL) framework that leverages local structural information and dynamic contextual relationships in the feature space. Our approach introduces a Dynamic Context Aggregation (DCA) mechanism, which integrates neighborhood information into sample representations, implicitly forming a dynamic graph structure. We further incorporate structure preservation constraints to maintain the intrinsic relationships of the data during contrastive learning. This method enhances feature representation quality and improves the model's ability to discover and represent unknown categories. Extensive experiments on generic, fine-grained, and handwritten character recognition tasks show that our CACL framework significantly outperforms state-of-the-art GCD algorithms, showcasing its effectiveness in handling complex, real-world data distributions and discovering novel categories. Notably, CACL achieves comparable performance even without prior knowledge of the actual number of classes, demonstrating its robustness in realistic scenarios.