Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.794658
Title: Visibility metrics and their applications in visually lossless image compression
Author: Ye, Nanyang
ISNI:       0000 0004 8500 4931
Awarding Body: University of Cambridge
Current Institution: University of Cambridge
Date of Award: 2020
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Abstract:
Visibility metrics are image metrics that predict the probability that a human observer can detect differences between a pair of images. These metrics can provide localized information in the form of visibility maps, in which each value represents a probability of detection. An important application of the visibility metric is visually lossless image compression that aims at compressing a given image to the lowest fraction of bit per pixel while keeping the compression artifacts invisible at the same time. In previous works, most visibility metrics were modeled based on largely simplified assumptions and mathematical models of human visual systems. This approach generally fits well into experimental data measured with simple stimuli, such as Gabor patches. However, it cannot predict complex non-linear effects, such as contrast masking in natural images, particularly well. To predict visibility of image differences accurately, we collected the largest visibility dataset under fixed viewing conditions for calibrating existing visibility metrics and proposed a deep neural network-based visibility metric. We demonstrated in our experiments that the deep neural network-based visibility metric significantly outperformed existing visibility metrics. However, the deep neural network-based visibility metric cannot predict visibility under varying viewing conditions, such as display brightness and viewing distances that have great impacts on the visibility of distortions. To extend the deep neural network-based visibility metric to varying viewing conditions, we collected the largest visibility dataset under varying display brightness and viewing distances. We proposed incorporating white-box modules, in other words, luminance masking and viewing distance adaptation, into the black-box deep neural network, and we found that the combination of white-box modules and black-box deep neural networks could generalize our proposed visibility metric to varying viewing conditions. To demonstrate the application of our proposed deep neural network-based visibility metric to visually lossless image compression, we collected the visually lossless image compression dataset under fixed viewing conditions and significantly improved the deep neural network-based visibility metric's accuracy of predicting visually lossless image compression threshold by pre-training the visibility metric with a synthetic dataset generated by the state-of-the-art white-box visibility metric---HDR-VDP \cite{Mantiuk2011}. In a large-scale study of 1000 images, we found that with our improved visibility metric, we can save around 60\% to 70\% bits for visually lossless image compression encoding as compared to the default visually lossless quality level of 90. Because predicting image visibility and predicting image quality are closely related research topics, we also proposed a trained perceptually uniform transform for high dynamic range images and videos quality assessments by training a perceptual encoding function on a set of subjective quality assessment datasets. We have shown that when combining the trained perceptual encoding function with standard dynamic range image quality metrics, such as peak-signal-noise-ratio (PSNR), better performance was achieved compared to the untrained version.
Supervisor: Mantiuk, Rafal Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.794658  DOI:
Keywords: Visibility metrics ; Visually lossless image compression ; Deep learning
Share: