Estimating the usefulness of distorted natural images using an image contour degradation measure

Abstract

Quality estimators aspire to quantify the perceptual resemblance, but not the usefulness, of a distorted image when compared to a reference natural image. However, humans can successfully accomplish tasks (e.g., object identification) using visibly distorted images that are not necessarily of high quality. A suite of novel subjective experiments reveals that quality does not accurately predict utility (i.e., usefulness). Thus, even accurate quality estimators cannot accurately estimate utility. In the absence of utility estimators, leading quality estimators are assessed as both quality and utility estimators and dismantled to understand those image characteristics that distinguish utility from quality. A newly proposed utility estimator demonstrates that a measure of contour degradation is sufficient to accurately estimate utility and is argued to be compatible with shape-based theories of object perception.

Recognizable or Not: Towards Image Semantic Quality Assessment for Compression

“…For example, how to quantify the semantic quality of image for other image analysis tasks, including detection, tracking, classification, retrieval, and so on? What is the relation between visual quality, perceived utility [19], and semantic quality? Is the proposed ISQA measure convex, or how to design a convex ISQA measure for easy optimization?…”

“…Rouse et al propose to ask human subjects to evaluate the utility/usefulness of images, which is named utility assessment task in contrast to quality assessment task; they report that perceived quality and perceived utility are correlated but clearly different, a perceived quality score is not a proxy for a perceived utility score, and vice versa [19,20]. Their proposed utility assessment task is performed by human and objective utility assessment is designed to predict perceived utility, like objective IQA is to predict perceived quality.…”

“…But in our work, we consider the semantic quality of images directly from the performance of semantic analysis algorithms when working on the images, we do not involve perceived utility at all. The relationship between utility assessment in [19] and our proposed ISQA is to be investigated in future work.…”

Rate-perceptual-distortion optimization (RpDO) based picture coding — Issues and challenges

“…In constant visual quality coder design and services, however, a perceptual distortion measure must be able to detect JNND as the reference quality level at which the compressed picture is indistinguishable from the original, and predict consistently discernible levels by human visual perception in terms of JND 1 , JND 2 , etc. How this set of perceptually discernible levels of distortions/quality is mapped into practical scales of various applications, e.g., perceptual utility of pictures (UoP) [21], depends very much on the application in question and requires participation of targeted human observers who have necessary domain knowledge of intended applications, e.g., radiologists and radiographers in medical diagnostic imaging [6].…”

“…Rate-distortion optimization considering a perceptual distortion measure [1] or a utility score [21] for QoE regulated services compared with the MSE. In [21], RT (recognition threshold) is defined as a perceived utility score threshold with a value of zero (0) below which an image deems to be useless and REC (recognition equivalence class) defines a class of images whose perceived utility score with a value of 100 (i.e., REC (100)) is statistically equivalent to that of a perceptually lossless image with respect to and including the reference. Black solid line corresponds to R-D curve which can be optimized towards black dash line.…”