The role of perceptual contrast non-linearities in image transform quantization

Normalized image representation for efficient coding

“…The quantizers based on the minimization of the MSE end with non-uniform quantization solutions based on the marginal PDFs [4] or some modification of them including the perceptual metric [2], [14]. However, it has been suggested that constraining the Maximum Perceptual Error (MPE) may be better than minimizing its average [14].…”

“…The difference between the approaches that implicitly followed the MPE idea [2], [14]- [18] is the accuracy of the perception model which is used to propose the perceptually Euclidean domain. For instance, the quantization scheme (empirically) recommended in the JPEG standard [16] may be deduced from the MPE restriction with a very simple linear vision model based on the CSF [14].…”

“…For instance, the quantization scheme (empirically) recommended in the JPEG standard [16] may be deduced from the MPE restriction with a very simple linear vision model based on the CSF [14]. In this case, the perceptual metric is fixed (the model is linear) and it is assumed to be diagonal in the local DCT domain.…”

“…In this very simple case, it is assumed that no perceptual relationship exists between the coefficients of the transform, and that the perceptual relevance of each coefficient is given by the corresponding CSF value. The performance of this approach can be improved at around 0.5 bits/pix if a more sophisticated model is used [14], [15], [17], [18]. In these references the authors used a point-wise non-linear model in the DCT domain.…”

“…5, top-left), the algorithm of Malo et al [14], [18], that assumes a point-wise non-linear model [11] and hence an input-dependent diagonal metric in the DCT domain ( fig. 5, top-right), the algorithm of Epifanio et al [2], that assumes a fixed non-diagonal metric in the DCT domain ( fig.…”