Quality Evaluation of Tone Mapping HDR Omnidirectional Image with Multi-regions and Multi-levels

Unlike ordinary omnidirectional imaging, which produces underexposure or overexposure of some areas under wide range of lighting conditions, tone mapping high dynamic range omnidirectional image (denoted as TM-HOI) presents more detailed information with high dynamic range (HDR) techniques. Aiming at the multiple distortions in TM-HOI processing, this paper proposes a blind TM-HOI quality metric with multi-regions and multi-levels analysis. Considering the conversion of image projection formats and the unique distortion caused by user behavior in immersive environments, feature extraction in the proposed metric is divided into the viewport based module and the Crasher parabolic projection based module. At the same time, considering the coding distortion, tone mapping (TM) distortion and mixed distortion in the TM-HOI and the different manifestations of this mixed distortion in different regions, this paper further extracts perceptual features to characterize image distortion by performing bit plane layer decomposition and detail/basic layer decomposition on different regions of TM-HOI. Finally, the extracted perceptual features are used as the input of Random forest to construct the nonlinear relationship between the feature space and subjective opinion scores. The experimental results indicate that the proposed metric has better consistency with the human visual system compared to representative blind quality metrics.