Objective Image Quality Assessment Based Tone Mapping Optimization
Abstract
High dynamic range (HDR) images provide the capacity to represent the luminance in real scenes with much higher precision than standard image formats. With advances in hardware and computer graphics technologies, HDR images are rapidly becoming more commonplace. To visualize HDR images on contemporary display devices, the dynamic range needs to be adapted to the much smaller range of the devices. This is accomplished through tone mapping, with the goal of reproducing the visual appearance of HDR scenes. Tone mapping has attracted much attention and several dozens of tone mapping operators have been proposed.
Nevertheless, it remains challenging to objectively evaluate the quality of tone mapped images and optimize tone mapping operators with automated algorithms. Using virtual photographs to bridge the gap of dynamic ranges for feature analysis, we propose two feature-based quality metrics for tone mapped images, which measure the distortion of important image features that affect the perceived quality. We present an image quality metric called visual saliency distortion predictor (VSDP) that measures the distortion in visual saliency for quality assessment. Additionally, by incorporating multiple feature-based measures to predict the quality of tone mapped images, we introduce another quality metric: perceptual distortion predictor (PDP). Subjective and numerical experiments indicate that the proposed feature-based quality metrics can yield more reliable prediction than the alternative approaches.
Once suitable quality metrics are defined, there emerges an opportunity to automate the tuning of existing tone mapping operators. By minimizing the distortion in visual saliency predicted by the quality metric VSDP, we developed an automatic parameter tuning algorithm for tone mapping operators. Moreover, based on the quality prediction of PDP, we propose an automated blended tone mapping algorithm which blends images from multiple operators with varying weights to leverage the strengths of each of operators considered. Experiments with a broad range of HDR images and statistical analysis demonstrate the effectiveness of the tone mapping optimization algorithms.