Efficient Image Debiased Contrastive Clustering

The rise of e-commerce and social media has overwhelmed systems with image data, challenging real-time clustering and recommendation. Although multistage or large-pretrained-model (LPM) assisted clustering methods achieve high accuracy, they often suffer from large model sizes and high computational costs. Single-stage methods, while saving clustering resources, face challenges like weak augmentations limiting feature diversity and false-positives/negatives harming accuracy. We propose debiased contrastive clustering (DCC), an efficient lightweight model that addresses these issues by integrating differential augmentations, refined sampling, and debiased contrastive loss to reduce false negatives. Pseudo-labels and consistency regularization mitigate false positives, boosting accuracy without multistage training or LPM reliance. Experiments on seven challenging datasets show that DCC outperforms state-of-the-art (SOTA) methods in accuracy, normalized mutual information (NMI), adjusted Rand index (ARI), and efficiency, converging faster with superior results.