SegUnify: A Unified Framework for Unsupervised Semantic Segmentation

Semantic Segmentation is the process of assigning predefined classes to each pixel in an image. In Unsupervised Semantic Segmentation, no manual labels are provided, so the segmentation is performed by identifying patterns and correlations in the data. Earlier methods mainly used selfsupervised learning or clustering to achieve semantic coherence. However, those methods often fail to maintain consistency between different parts of an image and make errors while segmenting complex images. This research aims to solve these limitations by introducing SegUnify, an integrated framework for semantic segmentation on unlabelled images. The model outperformed benchmarks on the Potsdam-3 dataset, achieving 82.74% accuracy and a mean IoU of 70.34%, surpassing existing methods. These results demonstrate the ability of SegUnify to produce more coherent and accurate segmentation maps, contributing to advancements in unsupervised semantic segmentation.