A novel high-fidelity reversible data hiding scheme based on multi-classification pixel value ordering

Pixel value ordering (PVO) is a highly effective technique that employs a pixel block partitioning and sorting for reversible data hiding (RDH). However, its embedding performance is significantly impacted by block size. To address this, an improved pixel-based PVO (IPPVO) was developed adopting a per-pixel approach and adaptive context size. Nevertheless, IPPVO only considers pixels below and to the right for prediction, neglecting other closer neighboring regions, leading to inaccurate predictions. This study presents a RDH strategy using multi-classification embedding to enhance performance. First, pixels are categorized into four classes based on parity coordinates, obtaining higher correlation prediction values using an adaptive nearest neighbor content size. Second, a new complexity calculation method is introduced, the complexity frequency of pixel regions to better differentiate between complex and flat regions. Finally, an effective embedding ratio and index value constraint are introduced to mitigate the challenge of excessive distortion when embedding large capacities. Experimental results indicate that the proposed scheme offers superior embedding capacity with low distortion compared to state-of-the-art PVO-based RDH methods.