A Robust No-Reference, No-Parameter, Transform Domain Image Quality Metric for Evaluating the Quality of Color Images

Abstract

In autonomous imaging and video systems, data measurements can be extracted based on the presence or absence of system specific attributes of interest. These measurements may then be used to make critical system decisions. Therefore, it is imperative that the quality of the image used for extracting important measurements is of the highest fidelity. To achieve this, image enhancement algorithms are used to improve the quality of the image as a preprocessing procedure. Currently, most image enhancement processes require parameter selection and parameter optimizations, where the results typically require assessment by a human observer. To perform the image enhancement without human intervention, an image quality metric needs to be used to automatically optimize the enhancement algorithm’s parameters. An additional complexity is that the performance of an image quality measure depends on the attributes an image possesses and the types of distortions affecting the image. Although there are many image quality metrics available in the literature, very few are designed for color images. Furthermore, most color image quality measures require a reference image as a basis, on which all other results are compared too, or require parameter adjustment before the measures can be used. Finally, most available measures can only evaluate the image quality for images that are affected by a small set of distortions. In this paper, we will show a new no-reference no-parameter transform-domain image quality metric, TDMEC, which can successfully evaluate images that are affected by ten different distortion types in the TID2008 image database. This measure enables vision-based measurement systems to automatically select optimal operating parameters that will produce the best quality images for analysis.