No-reference visual quality assessment based on PSNR and ssim estimation for H.264/AVC-encoded video

Abstract

As the demand for high-quality video services in tandem with mobile video services increases, the needs for assessing the visual quality has also been increased. Therefore, visual quality assessment (VQA) monitoring techniques by service providers or consumers become an important issue. No-reference VQA methods without original video data allow service providers to efficiently manage the QoS for the vide con-tents being delivered. Mean squared error (MSE) based peak signal-to-noise ratio (PSNR) has widely been used as an objective quality assessment metric. In addition, Structural Similarity Metric (SSIM) has also been known as a good visual quality metric that properly reflects the human visual system (HVS). These quality assessment metrics can be used without original video by no-reference based quality assessment methods. The no-reference VQA methods can be categorized into three kinds, depending in the availability of coding related information or decoded videos: (i) utilizing only decoded videos measures the visual quality with pixel intensities; (ii) utilizing bitstreams only measures the visual quality with coding related data, e.g. QP, bitrate, transform coefficients, etc.; (iii) utilizing bitstreams plus decoded videos that is a so-called hybrid method use both coding related information or decoded videos. This dissertation first studies a no-reference method for estimating MSE that is a core part of SSIM and then an SSIM estimation method in quantized transform domain of H.264/AVC bitstreams. Since, the estimated MSE without the original reference video can be ex-pressed with the expectation on the squared difference between DCT coefficients before quantization and those after quantization, the distribution of DCT coefficients before quantization should be known a priori. To cope with this, we provides a solution in two ways: (i) a method to find proper weight vector sets that help to express the distributions for DCT coefficients before quantization...