An Exclusion-Reference Image Quality Dataset with Color and Spatial Aspects

High image quality is desired for all imaging devices. Large efforts have been made to accumulate data by human observers and to develop models to fit the visual results. However, most of the earlier works applied the images include color and spatial domains but the visual results were only reported as total image quality, which was found insufficient to build a model. With this in mind, the present experiment studied 816 images collected from two parts, 585 images rendered by color and spatial domain functions from two earlier published datasets (CIDIQ and KADID) and 231 images rendered by color domain functions from the authors’ earlier dataset. The present experiment was conducted using the exclusion-reference (ER) method. Thirty participants took part to evaluate 4 terms: total image quality (tIQ), color image quality (cIQ) and spatial image quality (sIQ) by means of a six categorical judgment method, as well as the weight between color impact and spatial impact (sum to 1). The results showed good repeatability performance between present and earlier visual score were 0.90 and 0.82 correlation coefficient values for CIDIQ and KADID respectively. In addition, weighted IQ obtained through linear weighting with cIQ, sIQ and weight had a greatly high correlation coefficient with tIQ of 0.96, which implies the feasibility of separately considering the color and spatial aspects of image quality. Furthermore, a no-reference images quality model was proposed to predict tIQ, whose accuracy of prediction obtained a correlation coefficient value of 0.80 with 80–20 train-test method.