A Novel Just Noticeable Difference Model via Orientation Regularity in DCT Domain

Just noticeable difference (JND) model plays an important role in removing perceptual redundancies for image/video compression. However, the existing subband-based JND models have two limitations: one is the evaluation of contrast masking (CM) effect is not comprehensive; the other is that the operation within cross domain is computational complexity. In this paper, we propose a new orientation regularity-based JND model to solve these problems in the discrete cosine transform (DCT) domain. Inspired by the structure regularity extraction in the human brain, we suggest that orientation features play an important role in the analysis of visual content. By deducing the DCT function, the orientation information is first analyzed with the DCT coefficients along different directions, and the orientation regularity is calculated with the coefficient distribution. Then, according to the orientation regularity and the frequency texture energy, the DCT blocks are classified into five types, i.e., smooth, orderly edge, disorderly edge, orderly texture, and disorderly texture. By combining these block types with their human visual system sensitivities, a more accurate CM model is proposed in a DCT domain. Finally, by incorporating the contrast sensitivity function and luminance adaptation effect, a novel DCT-based JND model is established. Since the proposed model introduces a more accurate CM model and performs only in a DCT domain, it is more efficient and concise than the state-of-the-art DCT-based JND models in theory and practice. The experimental results also show that the proposed JND model has a superior performance without cross domain.