Unsupervised Image Segmentation through Contrastive Learning and Graph-Based Clustering

Unsupervised image segmentation is a fundamental problem in computer vision, particularly for applications where labeled data are scarce or expensive to obtain. This paper presents a novel unsupervised image segmentation framework based on contrastive representation learning and graph-based clustering, designed to achieve high segmentation accuracy without relying on manual annotations. The proposed method first learns discriminative pixel- and region-level embeddings using a self-supervised contrastive learning strategy that maximizes intra-region similarity while enhancing inter-region separability. These learned embeddings are then modeled as a graph, where nodes represent image regions and edges capture feature similarity, enabling robust partitioning through graph-based clustering. Experimental evaluation is conducted on standard benchmark datasets, including BSDS500, Pascal VOC (unsupervised setting), and MSRC, to assess segmentation quality. The proposed approach achieves an average Intersection-over-Union (IoU) of 72.4%, pixel accuracy of 89.1%, and boundary F-score of 0.78, outperforming traditional graph-cut and recent unsupervised deep clustering methods by 6–11% across key metrics. Ablation studies further demonstrate that contrastive feature learning improves clustering compactness by approximately 18%, leading to more coherent and semantically meaningful segments. These results confirm that the integration of contrastive learning with graph-based clustering provides a powerful and scalable solution for unsupervised image segmentation, making the proposed framework well suited for large-scale image analysis in medical imaging, remote sensing, and autonomous vision systems.