No-reference image quality metrics for color domain modified images

Abstract

Predicting the quality of natural images without using a reference image has always been a challenging task. Numerous approaches have been proposed in the past, but they mainly focused on spatial and frequency domain degradations like blur, noise, and compression. Image quality metrics (IQMs) in literature perform with quite a high accuracy for such types of degraded images. However, their performances are not good on the images modified in the color domain. In this study, psychophysical experiments were conducted to assess the quality of the color domain images. A new dataset was developed for this purpose. Additionally, a second dataset consisting of color domain modified images from the three previously published datasets were used in the psychophysical experiments. The newly developed dataset was then used to develop three IQMs based on absolute values, relative values, and statistical analysis of image color appearance attributes. Their performances were then evaluated together with five spatial domain IQMs from the literature using cross-database evaluation methodology. The results showed that the color-domain IQMs outperformed the other models. The absolute and relative attributes-based models, when combined, achieved the best performance. The present results suggest that more effort is needed to improve the performance of color domain IQMs for image quality estimation.

A Comprehensive Image Quality Dataset to Compare No-reference and With-reference Image Quality Assessment

“…( 10). 𝑤𝐼𝑄 = 𝑐𝐼𝑄 * 𝑟𝑎𝑡𝑖𝑜 color + 𝑠𝐼𝑄 * 𝑟𝑎𝑡𝑖𝑜 spatial (10) Table 2 lists R between IQs and figure 6 plots the scatter between tIQ and other IQs. From the table and figure 6 (c), the correlation between cIQ and sIQ is close to 0 at 0.092, which indicated that it is feasible to consider image quality in terms of color and spatial quality.…”

“…For certain attributes, the computational procedure is to increase this attribute of the target pixel until it reaches the gamut boundary [10]. Then the end point on the gamut could be regarded as a reference point showed in figure 7.…”

“…At present, several image datasets have been developed and applied extensively in image quality modeling, such as the LIVE database of the University of Texas [5], the TID2008 and TID2013 datasets of the Finnish University of Tampere, the CSIQ dataset of the Oklahoma State University [6,7], the CIDIQ dataset of the Norwegian University of Science and Technology [8] and the KADID dataset [9] of the University of Konstanz in Germany. Recently, a dataset was developed at the Color and Engineering Lab of Zhejiang University consisting of 47 original images and whole 1600 images rendered by over 6 color domain modifications [10].…”