Reduced-Reference Methods for Measuring Quality Attributes of Natural Images in Imaging Systems - Part II
Mikko Nuutinen · 2012
This article is Part 1 of two parts summarizing the work presented in the doctoral thesis in media technology titled “Reduced-reference methods for measuring quality attributes of natural images in imaging systems”. This part includes a literature survey of image quality measurement methods and algorithms. Part II presents novel methods and algorithms for predicting subjective image quality attributes of natural images captured by cameras or printed by a printer, shows performance of the proposed methods and makes conclusions of the study. The dissertation is based on six publications, listed in the Publication of the Dissertation, and are referred to in the text by their Roman numerals. The whole dissertation is digitally available at http://lib.tkk.fi/Diss/2012/isbn9789526048819/isbn9789526048819.pdf. The main goal of the research field of image quality is to create a computational model capable of predicting the subjective visual quality of natural images and video. The model can be an alternative for expensive quality evaluations by human assessors. This study aligns with this research field in the case of imaging systems, such as cameras, displays and printers. The traditional approach to measuring the quality of imaging systems is based on test targets. These targets primarily facilitate description of the performance of a system in terms of how it reproduces and distorts simple test signals rather than measure the visual quality of natural images captured or shown by an imaging system. This study primarily contributes novel methods and algorithms for measuring the image quality attributes of natural images captured by cameras or printed by a printer. Both methods utilize reference image data in the reduced-reference mode. The method and algorithms developed for printers transform the printed natural test images into the form of the reference image by using a high-quality reference camera and multiple exposures. The methods and algorithms developed for camera images use a reference camera to capture scene information. The scene information is used to help measure the attributes of natural images. The main problem which needed to be solved concerns localization of areas in images from which different attributes can be measured. The challenge arises from a multidimensional distortion space in capture and display. The solution relies on low-level computational understanding of images. The methods were evaluated with subjective data. Compared with the state-of-the-art computational or test target metrics, these methods were highly effective at predicting the quality attributes of natural images captured by different cameras or printed on different papers. According to the results, the proposed methods can replace test target methods and even small-scale subjective tests in some situations.