Self-similarity based structural regularity for just noticeable difference estimation

In this paper, we introduce a novel just noticeable difference (JND) threshold estimation model based on a spatial masking function taking both luminance difference and structural regularity into account. Existing spatial masking functions underestimate the JND threshold for irregular textural regions, because they mainly consider the amplitude of luminance change for simplicity. As regular areas show weak masking effect due to their self-similar structures while irregular regions present strong masking effect, the spatial structure directly determines spatial masking. To effectively measure structural regularity in images under different contents, we propose an adaptive non-local self-similarity analysis based procedure. Then we weight luminance differences with similarity coefficients and deduce a new spatial masking function. Finally, an accurate JND estimation model is introduced. Experimental results demonstrate that the proposed JND model has a better visual effect than other models: it injects much noise into the insensitive regions, whereas little into the sensitive regions.

► We show that structural regularity is another determination of spatial masking.
► A non-local procedure is introduced to measure the structural regularity.
► A novel spatial masking function is derived based on the structural regularity.
► We propose a precise spatial domain just-noticeable difference model.