No-Reference Quality Assessment for HDR Images Based on Multi-scale Retinex Decomposition

High dynamic range (HDR) content is increasingly entering our life for super realistic representation of real-word scenes. Compared to low dynamic (LDR) images, HDR images contain significantly higher peak luminance and contrast, making traditional LDR image quality assessment (IQA) metrics inefficient for HDR image applications. Towards developing HDR IQA, we propose a new no-reference metric based on Retinex decomposition. First, multi-scale Retinex decomposition is used to generate luminance and reflectance maps through gaussian filtering. Then, we calculate gradient similarities and natural scene statistics (NSS) from the reflectance and luminance maps, respectively. Finally, all features are fused by support vector regression (SVR) for score prediction. Experimental results on two databases demonstrate the superiority of the proposed method over the other state-of-the-arts for HDR IQA.