Context‐Aware Dual Pixel Value Ordering: A Variance‐Driven Framework for High‐Fidelity Reversible Data Hiding

Reversible data hiding (RDH) based on pixel value ordering (PVO) attains high fidelity, yet most variants still rely on uniform region selection with fixed or heuristic thresholds, and apply embedding indiscriminately across the image blocks. This leaves a clear gap: high‐variance areas are modified even when the required capacity is small, and the embedding region cannot be adjusted in a streamlined manner to meet a specified capacity requirement. We propose a context‐aware dual PVO (CA‐DPVO) framework with a statistically grounded, variance‐driven eligibility model and an explicit payload‐aware control as a technical extension of our previously proposed DPVO. The image is partitioned into non‐overlapping blocks; for each block a local variance is computed, and a global threshold , set as the ‐quantile of the variance distribution, classifies blocks as eligible or ineligible for embedding. By dynamically tuning to the target capacity, CA‐DPVO confines modifications to low‐variance regions and adjusts the number of embedded blocks accordingly. Within the selected regions, a forward PVO pass embeds in the extremal pixels of the eligible blocks, and a backward pass forms min–max pairs to embed additional bits, while restoring earlier shifts. Extraction mirrors these stages and remains fully reversible. Notably, when , CA‐DPVO reduces seamlessly to the original DPVO, demonstrating that the proposed framework generalizes DPVO as a special case. Experiments on standard test images show consistent improvements over recent PVO‐based baselines, yielding higher image fidelity at matched capacity, while preserving strict reversibility.