OPTIMIZING IMAGE SEGMENTATION ACCURACY USING HYBRID CLUSTERING TECHNIQUES: A MACHINE LEARNING APPROACH

Abstract

Image segmentation is a fundamental task within computer vision and machine learning, allowing images to be effectively analysed so that they can be applied to remote sensing, urban analysis, and autonomous systems. Although there has been a large amount of research performed on clustering-based segmentation techniques, these methods can be limited by several issues, such as the sensitivity of the initial cluster assignment, noise in the dataset, and inaccurate boundary definition in complex scenes. To address these challenges, this study proposes a hybrid clustering-based image segmentation framework that integrates classical clustering techniques with machine learning-based feature refinement.

The approach combines K-means and Fuzzy C-means clustering to leverage both hard and soft segmentation capabilities. The approach then applies learning-driven refinements to increase boundary accuracy/robustness. Next, experiments were conducted using an accessible public aerial image dataset of Dubai, which was labelled at the pixel level for six land-cover classes. Segmentation performance was measured using Accuracy, Precision, Recall, F1-Score, Dice Coefficient, and Intersection over Union (IoU). The results demonstrated that this proposed method yielded 0.87 accuracy, 0.85 Dice coefficient, and 0.74 IoU when using hybrid instead of individual clustering methods. This shows how the hybrid framework provides an effective, interpretable, and computationally efficient method of segmenting complex aerial imagery and is particularly useful for segmenting aerial images when operating in data-constrained remote sensing environments.