Measuring contour degradation in natural image utility assessment: methods and analysis

Utility estimators predict the usefulness or utility of a distorted natural image when used as a surrogate for a reference image. They differ from quality estimators in that they should provide accurate estimates even when images are extremely visibly distorted relative to the original, yet are still sufficient for the task. Our group has previously proposed the Natural Image Contour Evaluation (NICE) utility estimator. NICE estimates perceived utility by comparing morphologically dilated binary edge maps of the reference and distorted images using the Hamming distance. This paper investigates perceptually inspired approaches to evaluating the degradation of image contours in natural images for utility estimation. First, the distance transform is evaluated as an alternative to the Hamming distance measure in NICE. Second, we introduce the image contour fidelity (ICF) computational model that is compatible with any block-based quality estimator. The ICF pools weighted fidelity degradations across image blocks with weights based on the local contour strength of an image block, and allows quality estimators to be repurposed as utility estimators. The performances of these approaches were evaluated on the CU-Nantes and CU-ObserverCentric databases, which provide perceived utility scores for a collection of distorted images. While the distance transform provides an improvement over the Hamming distance, the ICF model shows greater promise. The performances of common fidelity estimators for utility estimation are substantially improved when they are used in ICF computational model. This suggests that the utility estimation problem can be recast as a problem of fidelity estimation on image contours.