Announcing the 9th SPRINGEROPEN EURASIP JIVP's Free Web conferencing (Sep. 2, 2021)

27 Août 2021

---

Catégorie : Autres événements

---

Abstract: Three-dimensional video systems such as 3D Television and Free-viewpoint Television provide viewers with visual experience of depth perception and interactivity. 3D video applications need a large number of views at different viewpoints. For limited number of viewpoints, 3D content is captured by cameras. For other viewpoints, new video sequences are synthesized from captured video and from their associated depth maps using Depth-Image-Based-Rendering (DIBR) techniques. DIBR techniques introduce specific distortions mainly in the disoccluded areas (image regions that become visible from the position of the synthesized view, while invisible in the reference view) affecting the edge coherency that are non-uniformly spread over the image. Distortions in synthesized images also appear due to the distortions of texture images and depth maps, which may occur during the process of acquisition, compression and transmission. The dominant distortions in the synthesized images, annoying for humans, are edge distortion, distorted shapes, inconsistent object shifting and flickering in the synthesized video. The quality assessment of the synthesized view is important for the development of view synthesis algorithms and 3D imaging applications. Quality prediction of synthesized view can be used for rate/distortion optimization in compression algorithms.

In this talk, several metrics for quality prediction of synthesized views that we’ve designed will be presented. These metrics have been developed considering multi-resolution and multi-scale properties of human visual system when perceiving the visual scene. The metrics use morphological filters to maintain important geometric information such as edges and contours across different resolution levels and scales. First, reference-based metrics for quality assessment of DIBR-synthesized images, MP-PSNR based on morphological pyramid decomposition and MW-PSNR based on morphological wavelet decomposition, described in the EURASIP JIVP 2016 paper will be presented. Then, no-reference metrics using morphological wavelet decomposition, that we’ve designed later, NR-MWT and MW-GRNN, will be presented. Referenceless metrics are more desirable since reference views are often not available. Computationally highly efficient completely blind NR-MWT metric has been proposed to predict the quality of synthesized video with synthesis distortions only. Learning-based MW-GRNN metric using sparsity based features has been designed to predict the quality of synthesized views with high reliability.

Short bio: Dragana Sandić-Stanković received her Diploma degree in electronics in 1991, from the High Military-Technical School of Land Force, SFR Yugoslavia. She received MSc degree from the Faculty of Electrical Engineering, University of Belgrade with thesis "Morphological filters for multiresolution decomposition and coding of monochrome images" and PhD degree in electrical and computer engineering from the Faculty of Technical Sciences, University of Novi Sad, Serbia with thesis "Multi-resolution metric for Objective Quality Assessment of Synthesized Views for Free Viewpoint TV". She has been working on the software development for Digital Switching System for Serbian Telecommunication Industry. She is currently a researcher at the Institute for Telecommunications and Electronics, IRITEL, Belgrade. She has authored over 50 scientific and technical publications. Her research interests include the applications of multiresolution and multiscale decompositions using morphological filters in image and video processing, image and video quality assessment.