Structural uncertainty based just noticeable difference estimation

Just noticeable difference (JND) reveals the minimum visible threshold of the human visual system (HVS), which is useful in visual redundancy reduction. Existing JND models estimate the visible threshold with luminance adaptation and contrast masking. As a result, the smooth and edge regions are effectively estimated, while the disorderly texture regions are always underestimated. The disorderly texture regions possess a large amount of disorderly structures and the HVS cannot fully perceive them. Therefore, in this work, we suggest to consider the disorder degree of structure for JND threshold estimation. According to the correlation among neighboring pixels, the uncertain information is extracted, and the disorder degree of structure is computed, which we called structural uncertainty. Then, taking the effect of background luminance, contrast, and structural uncertainty into account, a novel JND model is deduced. Experimental results demonstrate that the proposed JND can accurately estimate the visible thresholds of different image regions. Moreover, the proposed JND is adopted to remove visual redundancy for JPEG compression, which saves about 14% bit rate while keeping the perceptual quality.