Abstract
Perceived visual quality in stereoscopic videos incorporates attributes from production, compression, transmission and rendering in the end-user terminal. By estimating perceived video quality, network resources can be optimised while a specific level of user satisfaction is achieved. Common to all prior research is the observation that perceived quality depends on quality of service (QoS) parameters during aforementioned stages. However, stereoscopic videos under same QoS conditions may have different perceived visual quality due to different content types. In this thesis, perceived visual quality, which is measured via subjective experiments, is estimated for different operation modes: offline, live and real-time. Offline estimation involves in a contributed contextual feature extraction for 3-D videos to differentiate video types in several categories. Content properties are spatio-temporal activity analysis for dominant depth layers. For offline estimation of subjective quality, several machine-learning algorithm like neural networks (NN), Gaussian mixture models (GMM) and linear discriminant analysis (LDA) are employed. For live and real-time tasks, it is important to cut dependency on reference video, as availability of reference signal to receiver side is not applicable. This is the main inspiration to propose an objective quality metric exclusively for 3-D stereoscopic video. The proposed metric is reduced reference (RR). The proposed metric is a measurement of overall 3-D quality containing information from texture information and edge properties for colour and depth sections of 3-D video. Grey level co-occurrence matrices (GLCM) are considered to model texture information for colour and depth sections. Moreover, colour and depth characteristics are combined unevenly to weight image quality and depth perception. The proposed metric has a very high correlation coefficient (CC) to opinion score for impaired 3-D videos with both compression and transmission artefacts. The proposed metric is used to estimate subjective quality in live and real-time modes applying decision tree (DT) technique. Output of live (post-encoder) technique can be fed into compression block to block further degradations of visual quality. The proposed realtime estimation is highly accurate to predict subjective quality of 3-D videos in the presence of channel effects. A feedback from this block can be sent to transmission units to correct effective parameters and avoid quality degradations.