Dr Ahmed Hassan is a postdoctoral fellow who hails from Sudan. His project focused on video quality assessment. He assessed a system to measure the quality of video streaming. Video streaming entails several processing steps, any of which can degrade the quality of a video. Being able to measure the visual quality of a video is essential to obtain the leading edge in many sectors of society. This can be done through video multimethod assessment fusion. The ML system learns from human perceptions of quality by imitating human preferences, which are configured to obtain optimal performance.

WHY THE NEED FOR VIDEO QUALITY ASSESSMENT?

Video quality assessment models play a key role in the video industry. The increasing demand for videos in security, entertainment and communication has generated an interest in video-related technologies. Video streaming involves several processing steps, including compression, scaling and transmission, that can result in visual artifacts in transmitted videos. Video quality assessment models measure the amount of visual degradation in processed videos.

VIDEO MULTIMETHOD ASSESSMENT FUSION (VMAF)

VMAF is the state-of-the-art data-driven video quality assessment developed by Netflix. It correlates well with the human judgement of video quality by combining several elementary metrics based on the premise that multiple metrics compensate for the weaknesses of individual metrics. These elementary metrics are image-based metrics, which capture the spatial information in the video (images) and the temporal information (motion). These metrics include visual information fidelity, detail loss metric and mean co-located pixel difference.

GENETIC ALGORITHMS

Genetic algorithms proved to be effective for hard problems, especially for automated design. They belong to a class of computational intelligence techniques called evolutionary algorithms, which are inspired by natural selection. Genetic algorithms solve a problem by evolving a population of candidate solutions using biologically inspired operators (mutation, crossover and selection).

THE AUTOMATED DESIGN OF VMAF

The design of VMAF is challenging as it involves decisions that are not straightforward to make. Moreover, it is time-consuming. Although VMAF is well designed by Netflix experts, the configuration space of VMAF is too vast to be explored manually. Ahmed proposed automating the design of VMAF using a genetic algorithm by considering three design decisions:

• The choice of elementary metrics that make up the basic components of VMAF (more than 40 metrics were used)
• The values of hyperparameters, which are model variables that affect performance
• The choice of aggregation methods that combine the scores computed per video frame

RESULTS

The human rating of visual quality was compared to the automated design. This approach was found to agree with the human ratings on the Netflix dataset. The automated design was consequently found to outperform the latest VMAF release in two large video datasets.

PROJECT HIGHLIGHS

Video quality assessment (VQA) models play a key role in the video industry. Netflix has recently developed a VQA model (VMAF), which correlates well with the human judgement of video quality. VMAF combines elementary metrics that capture the spatial information in the video (images) and the temporal information (motion). The design of VMAF is challenging as it involves decisions that are not straightforward to make. Moreover, it is time-consuming. It is proposed to automate the design of VMAF using a genetic algorithm. The automated design outperforms the latest VMAF release in two large video datasets. This study is valuable since it produces better VQA models with minimal human intervention.