Multi-Scale Structure-Guided Region-of-Interest Extraction Without Semantic Prior

This paper presents a deterministic, multi-scale method for extracting regions of interest (ROIs) based exclusively on local structural organization, without relying on semantic priors, object detection models, or learned representations.

The proposed approach builds upon a dense structural map (Φ), computed locally over sliding windows, and introduces a multi-scale consensus mechanism to reduce sensitivity to scale selection. Structural responses are evaluated across multiple spatial scales (ws ∈ {5, 7, 9, 11}), and a majority-vote strategy is used to retain only regions consistently selected across scales. This process suppresses unstable activations while preserving spatially coherent structures.

The method is fully deterministic and interpretable. It does not require training data, semantic labeling, or domain adaptation. ROI extraction is performed through structural thresholding, non-maximum suppression, and spatial constraints, with explicit control over maximum ROI count and coverage.

Extensive evaluation is conducted on 150 natural images. The study quantifies ROI coverage and spatial reduction, regional structural contrast metrics, inter-scale stability using non-parametric statistical tests (Wilcoxon signed-rank and Friedman tests), and Kendall’s W coefficients to assess consistency across window sizes.

Results show that the multi-scale consensus significantly reduces sensitivity to window size selection while producing more stable and coherent ROIs. Although the multi-scale configuration increases overall ROI coverage compared with the most selective mono-scale setting, it improves inter-scale consistency and reduces fragmentation effects.

The proposed framework provides a scale-robust structural basis for downstream image processing tasks such as selective enhancement and structure-constrained super-resolution. It shows that reliable ROI extraction can be achieved through structural coherence alone, without semantic modeling.