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

A multi-scale formulation of a local structural measure for region-of-interest extraction. Structural responses are computed at several window sizes and combined by inter-scale voting. On 150 images:- **Mask coverage:** multi-scale masks cover more of the image than single-scale ones (32–33 % against 18–19 %).- **ROI coverage:** 20–23 % against 4–8 %.- **Number of ROIs:** fewer (median 3–4). Most of this difference already appears with a single scale, where no vote takes place, so it cannot be attributed to the consensus itself. The multi-scale mask does not depend on any single window size. Its sensitivity to the choice of scales and to the vote threshold was not measured. The local measure selects the same regions as a local Shannon entropy map (paired median difference 0.0000). **Version notes (rev3)** - Coverage at K = 1 (no vote) shows that the gap is not due to the consensus; the text now says so.- A coefficient-of-variation comparison computed on different axes was removed.- The masks are described as more extended, not more compact.- A table that was empty in the previous version was restored from the measured data.- The reference list was reduced to the works actually cited.