Mutual Neighborhood-Enhanced Protypical Contrastive Learning for Clustering

Deep clustering has achieved remarkable progress due to the strong representation learning capabilities of deep neural networks. However, many popular deep clustering algorithms based on contrastive or non-contrastive strategies focus solely on instance-level or cluster-level representation learning, lacking the ability to explore cross-level information necessary for effective clustering. Moreover, these algorithms often overlook the semantic neighborhood information around instances or clusters, which may lead to suboptimal performance. To address these issues, we propose a novel and efficient end-to-end deep clustering framework termed Mutual Neighborhood-Enhanced Representation Learning for Clustering (MNC), which integrates Neighborhood-Enhanced Prototype-level Contrastive Learning (NPCL) with an Aligning with Consistency Samples (ACS) strategy. In particular, the network architecture comprises four main modules, namely, the feature extraction module, the consistency samples search module, the class-level contrastive module, and the instance-level non-contrastive module. Notably, we extend the conventional non-contrastive instance-level representation learning by incorporating operations on deeper neighborhood structures, leveraging mutual neighborhood relationships from dual views. Furthermore, we perform contrastive learning at the cluster-level based on the compact neighborhood structure in a spherical feature space, which encourages intra-cluster compactness and inter-cluster separability. Extensive experiments on five challenging image benchmarks demonstrate the superiority of our approach over the state-of-the-art. In particular, our approach achieves an ACC of 0.342 (0.804) on the Tiny-ImageNet (STL-10) dataset, which surpasses the second-best method by 7.9% and 4.5%, respectively. The code is available at https://github.com/SandLYJ.