No-reference quality assessment of tone-mapped image using clustering perception and salient regions detection

Abstract

In order to display a high dynamic range (HDR) image on a standard display device, we need to compress an HDR image into a tone-mapped image (TMI) using the tone-mapped operator (TMO). In order to evaluate the performance of different TMOs, a blind quality evaluation method of TMI is proposed. First, the TMI is divided into bright, dark and middle regions by clustering algorithm, and the Y channel is extracted in these three regions respectively to calculate the goodness of Gaussian fitness of the histogram. Second, the TMI is decomposed by tucker decomposition, and the texture features are extracted on the first sub-band. After that, the non-negative matrix is applied to calculate the salient regions of TMI, and features in the salient regions are extracted. Then, for the sake of comprehensive evaluation the quality of TMI, the naturalness feature is extracted in the gray scale domain and colorfulness features are extracted in HSI color space. Finally, the final score quality is predicted using support vector regression model training. Experimental results on TMID and ESPL-LIVE HDR database proved the proposed method is very effective and superior to the existing blind TMI quality assessment methods.