Global Distribution Adjustment and Nonlinear Feature Transformation for Automatic Colorization

Figure 5

The distribution of learned feature values (after E-step) is not Gaussian (a). After nonlinear transformation in M-step, this distribution becomes more Gaussian as plotted in (b) that can help increase the accuracy of classifying feature vectors leading to the more precision in color reconstruction. As shown in (b), the defects are removed in recovered image.

(a)

(b)