Sparsity based Video Quality Assessment

Muhammed, Shabeer P and Channappayya, Sumohana (2017) Sparsity based Video Quality Assessment. Masters thesis, Indian Institute of Technology Hyderabad.

Text
EE15MTECH11008.pdf - Submitted Version
Restricted to Registered users only until 11 July 2020.
Download (2MB) | Request a copy

Abstract

In my research work, I developed a spatio-temporal objective quality algorithm to assess the perceptual quality of natural videos in the Full-Reference (FR) as well as No-Reference (NR) settings. A common thread running through a majority of Video Quality Assessment (VQA) algorithm is that spatio-temporal quality is evaluated by pooling spatial and temporal quality that is measured independently. The proposed method is holistic in the sense that we evaluate spatio-temporal quality using sparse spatio-temporal representations. In other words, we do not pool independently measured spatial and temporal quality, but rather, jointly perform the spatio-temporal quality assessment. In this thesis, I present a novel Full-Reference (FR) video quality assessment (VQA) algorithm that op- erates on the sparse representation coe ffi cients of local spatio-temporal (video) volumes. Proposed work is motivated by the observation that the primary visual cortex of human visual system adopts a sparse coding strategy to represent visual stimulus. We rely on the K-Singular Value Decomposition (K-SVD) algorithm to construct spatio-temporal dictionaries for sparsely representing local spatio-temporal volumes of natural videos. Then show that the sparse representation coe ffi cients of these local volumes possess excellent dis- tortion discrimination properties. A family of dictionaries that span a wide range of spatio-temporal scales is constructed. The distance between the sparse coe ffi cients of the reference and distorted videos over the family of dictionaries serves as the feature representing perceptual quality. Support Vector Regression (SVR) is used to learn the relation between this feature and Di ff erence Mean Opinion Score (DMOS). The proposed algorithm delivers competitive performance over the LIVE standard definition (SD), EPFL standard definition (SD) and the LIVE Mobile high definition (HD) databases. Furthur, I also propose No-Reference (NR) video quality assessment (VQA) algorithm that operates on the sparse representation coe ffi cients of local spatio-temporal volumes. Empirically demonstrate that the his- togram of the sparse representations corresponding to each atom in the dictionary can be modelled well using a Generalised Gaussian Distribution (GGD). Then show that the GGD model parameters are good features for distortion estimation. The GGD model parameters corresponding to each atom of the dictionary form the feature vector that is used to predict quality using Support Vector Regression (SVR). The proposed algorithm delivers competitive performance over the LIVE SD, EPFL SD, LIVE HD and CSIQ video databases. The proposed FR-VQA and NR-VQA algorithms are simple and computationally e ffi cient as compared to other state-of-the-art FR-VQA as well as NR-VQA algorithms.

[error in script]

IITH Creators: