Selective Image Enhancement Driven by Local Structural Information

This paper presents a selective image enhancement method guided by local structural information rather than semantic or learned priors. A multi-scale formulation of the structural entropy measure ϕ is introduced to correct the scale dependency of single-window estimators. Structural responses are aggregated across multiple spatial scales using an inter-scale consensus mechanism to guide local CLAHE enhancement. Experimental results show that multi-scale ϕ-guided CLAHE achieves the best overall trade-off among all evaluated variants, outperforming gradient- and entropy-based baselines in terms of enhancement strength, spatial confinement, and robustness. The approach is fully deterministic, non-parametric, and non-generative, making it suitable for safety-critical and embedded image processing pipelines.