Abstract
Humans can easily assess the unnaturalness of images with artificial contrast manipulations. Previous studies have shown that the deviation from natural scene statistics (NSS) of the luminance distribution in an image (mean, standard deviation, skewness, kurtosis, and entropy) can be a strong predictor for the perceived unnaturalness of contrast-distorted images (Fang et al., 2015; Gu et al., 2013). While these studies directly used the global NSS computed across the entire image, here we show that the human visual system can assess the contrast naturalness only from local image statics. We conducted a psychophysical experiment where the participants evaluated the naturalness of local patches (128 x 128 pixels) extracted from 114 contrast-distorted and 6 original color images taken from Kodak dataset. The participants also evaluated their global naturalness in a separate session after the local evaluations. The stimuli were generated by linear shifting (in the range of [-200, 190]) and scaling (in the range of [-0.2, 2.0]) of intensity values. The participants observed each test patch without referencing to its undistorted counterpart. The results showed that the perceived naturalness of local patches was mostly consistent with that of the global image they were extracted from, even though the local statistics did not necessarily match those of the global ones. Regression analysis revealed that the likelihood of local image statistics with respect to the local NSS distributions could predict both the local and global unnaturalness well, while a model trained on the global NSS was less predictive of the local unnaturalness. These results suggested that the visual system can exploit the NSS of local image regions for contrast naturalness perception, despite the fact that the local NSS has a broader range of distribution and is less reliable than global NSS.