Blind Tone-mapped Image Quality Assessment and Enhancement via Disentangled Representation Learning

For compatibility with existing low dynamic range (LDR) display devices, the tone mapping operator (TMO) is widely applied to the high dynamic range (HDR) image, which inevitably leads to visual quality degradation. Various blind image quality assessment models have been developed to quantity the distortion degrees across different HDR images. However, these models only extract the quality-aware features and serve as a selector for different TMOs, which excludes visual content information and fails to conduct an end-to-end image enhancement towards a desired quality score. In this paper, we propose to jointly conduct blind tone-mapped image quality assessment and enhancement via disentangled representation learning. An encoder is firstly used to map the input image into the general feature space. Then, two branches are separately developed to extract the quality-aware and content-aware latent representations from the general feature, which are supervised with the quality score and image reconstruction constraints, respectively. Meanwhile, these two branches are also coupled with the adaptive instance normalization, which enables our model to flexibly modify the image towards any desired quality score. Extensive experiments confirm the effectiveness of the proposed method.