Structure-Guided Region-of-Interest Extraction Without Semantic Priors

This work presents an unsupervised method for region-of-interest (ROI) extraction based on local informational structure, without semantic priors, annotations, or task-specific models. ROI selection is reformulated as a problem of spatial organization and effective information density rather than object recognition or perceptual saliency. The approach relies on dense structural maps derived from an effective information measure, which are spatially aggregated to extract compact and coherent regions of interest. The method is fully non-supervised, task-agnostic, and applicable to scenarios where scene content or downstream tasks are unknown. Experimental results on natural and aerial image datasets show that the proposed approach achieves strong spatial reduction while preserving structurally meaningful regions, outperforming classical gradient-based operators and exhibiting performance comparable to entropy-based methods. This work provides a generic, structure-guided building block for ROI selection in computer vision pipelines, particularly suited to real-time, embedded, and multi-task applications.