摘要 : After an image or video is acquired, it is subjected to a number of signal processing algorithms, such as editing, enhancement, compression, scaling, transmission, and rendering, all of which can alter the perceptual quality. Visu... 展开 After an image or video is acquired, it is subjected to a number of signal processing algorithms, such as editing, enhancement, compression, scaling, transmission, and rendering, all of which can alter the perceptual quality. Visual quality assessment (VQA) aims to find an estimate of the impact of such alterations to be used, e.g., during algorithm/system design, optimization and comparisons (which employ benchmarking), or in quality monitoring. A wide variety of VQA approaches have been developed that seek agreement with the response of the human visual system to visual signal stimuli, which is very difficult to model, and hence, quite a challenge to predict. In the particular case of benchmarking and monitoring SLAs, VQA can have a big impact on the product or service under study or the parties involved, hence, one should be very cautious in trying to model the accordance with the human perception. Moreover, it is important to take the disagreement and sometimes conflicting opinions of subjects into account. Integrating machine learning (ML) modules can lead to faster and more accurate objective quality measures, but the available ML algorithms are not necessarily suitable for the task of objective quality assessment. This paper gives examples of data-driven modules in the context of benchmarking and SLAs. To this end, we continue on the work of Hemami and Reibman in this area and elaborate in particular on VQA employing so-called probabilistic pairwise preference and objective quality thresholds to enable real-life visual quality benchmarking and monitoring. 收起