Abstract

Image clustering remains a fundamental challenge in computer vision and machine learning, with applications spanning content-based image retrieval, object recognition, and visual data organization. High-dimensional image data and semantic relationships between visual concepts are often not captured by traditional clustering methods. This research proposes a novel framework that integrates semantic feature mapping with spanning-tree based optimization techniques to achieve superior image clustering performance. Our approach leverages deep convolutional neural networks for extracting semantic features, followed by a minimum spanning tree construction that preserves local geometric structures while maintaining global consistency. The spanning tree serves as a backbone for hierarchical clustering, enabling efficient graph-based optimization through edge weight refinement. We introduce an adaptive distance metric that combines visual similarity with semantic coherence, addressing the semantic gap in traditional image clustering methods. The proposed algorithm employs a two-stage optimization process: first constructing an initial spanning tree based on semantic features, then iteratively refining cluster assignments through local neighbourhood analysis. Preliminary experiments on benchmark datasets including CIFAR-10, ImageNet-1K subset, and COCO demonstrate significant improvements over state-of-the-art methods, achieving 15-23% higher normalized mutual information scores and 18-27% better clustering accuracy. The framework also exhibits robust performance on imbalanced datasets and maintains computational efficiency with O(n log n) time complexity for n images. This research contributes both theoretical insights into semantic-geometric feature spaces and practical algorithms for large-scale image organization systems.