DLGC: Dynamic Label-Guided Clustering for Open-World Novel Class Discovery

Novel Class Discovery (NCD) transfers knowledge from labeled known classes to identify unseen classes in unlabeled data. Existing methods isolate novel class discovery while neglecting two critical issues: underutilized semantic guidance from known classes and rigid reliance on predefined cluster numbers, limiting adaptability in dynamic open-world scenarios. We propose DLGC, a unified framework that bridges this gap through two innovations: semantic-aware feature alignment, which constructs a shared embedding space by integrating known class labels as discriminative constraints to optimize cluster boundaries, and self-evolving adaptation, dynamically inferring optimal cluster numbers via density-driven iterative validation without predefined assumptions. Experiments on CIFAR-10, CIFAR-100, and ImageNet-100 demonstrate DLGC's superiority over GCD, ORCA, PrCAL and OpenNCD, achieving 4.5%, 20.2%, and 4.7% improvements in accuracy for novel classes while reducing cluster estimation errors by a maximum of 9%. This validates DLGC's ability to unify supervised and unsupervised learning in evolving data environments.