The impact of phase transition on quality assessment of natural images
Ning Liu, Guangtao Zhai · 2016
Recent research indicates that the well known scale invariance property of natural images may originate from the existence of the critical image layer exhibiting phase transition from disordered to ordered states as the intensity scale increases. In this paper we examine the impact of phase transition on perceptual quality assessment of images and confirm that using only the ordered or structured information in the image as identified by phase transition can lead to improved performance, i.e. “part is better than all”. To explicitly locate the phase transition layer, we collected 200,000 10×10 patches from the hierarchically decomposed 8-bit images from the Kodak database and trained a 2D Ising model. Analysis of the Ising model parameters suggests that phase transition happens approximately at the 4-th layer (4-th bitplane from the least significant bit). We have also tested the performance of image quality metrics using different combinations of the intensity layers. It is interesting to notice that for all the test scenarios, the combined 4-th to 8-th layers, corresponding to the ordered or structured image information give the best performance. This observation confirms that our visual perception does rely on structured information when performing visual quality assessment. And the structured information of images can be extracted by the hierarchical decomposition with a phase transition model to facilitate perceptual image quality assessment.