A perceptual preprocess method for 3D-HEVC

Abstract

A perceptual preprocessing method for 3D-HEVC coding is proposed in the paper. Firstly we proposed a new JND model, which accounts for luminance contrast masking effect, spatial masking effect, and temporal masking effect, saliency characteristic as well as depth information. We utilize spectral residual approach to obtain the saliency map and built a visual saliency factor based on saliency map. In order to distinguish the sensitivity of objects in different depth. We segment each texture frame into foreground and background by a automatic threshold selection algorithm using corresponding depth information, and then built a depth weighting factor. A JND modulation factor is built with a linear combined with visual saliency factor and depth weighting factor to adjust the JND threshold. Then, we applied the proposed JND model to 3D-HEVC for residual filtering and distortion coefficient processing. The filtering process is that the residual value will be set to zero if the JND threshold is greater than residual value, or directly subtract the JND threshold from residual value if JND threshold is less than residual value. Experiment results demonstrate that the proposed method can achieve average bit rate reduction of 15.11%, compared to the original coding scheme with HTM12.1, while maintains the same subjective quality.Keyword:3D high efficient video coding(3D-HEVC), just-noticeable-distortion(JND), depth information, saliency model

A Survey of Stereoscopic 3D Just Noticeable Difference Models

“…The MJNDD model combines STJND with an adapted JNDD model, which segment the texture frame into background regions (BR) and foreground regions (FR). Recently, Shi et al [75] developed a new 3D-JND model, which considers the depth information and visual saliency in addition to LA, CM, and TM.…”

“…This method validated the authenticity and integrity of stereoscopic images by localizing the tampered regions. MJND model has been used in order to improve the efficiency of 3D-HEVC [75] and MVD video coding [74]. SJND was used for SVQA, whereas HJND and DJND have been applied to improve 3D CE for MVD.…”