A Wavelet and Clustering Hybrid for Image Segmentation: A Probabilistic Rand Index Evaluation

Image segmentation refers to the process of partitioning a digital image into meaningful regions where homogeneity of attributes like color or texture exists among the pixels of the same region, and it forms an integral component of the workflows in computer vision, like in medical imaging, imaging satellites, or even image-based object detection and recognition. K-means and Fuzzy C-means, which are classical clustering techniques, are known to be quite fast in processing, but have issues with noisy or irregular borders and a fixed number of clusters, which is quite a common limitation in many fields. This paper proposes a spatial-contrastive adaptive clustering (SCAC) image segmentation approach, which combines self-supervised contrastive feature learning, adaptive density-based clustering, and spatial consistency refinement. The proposed framework includes a contrastive encoder that processes unlabeled images to derive features with rich semantics. The features are reduced to lower dimensions to improve the clustering results. The proposed approach is evaluated using probabilistic (PRI) on the benchmark BSDS300 data set. Experimental results indicate that SCAC outperforms conventional clustering techniques, achieving cleaner boundaries, greater noise robustness, and higher segmentation accuracy, thereby establishing it as an effective framework for unsupervised image segmentation.

• Supervision