ARTIFACT DETECTION IN GAMUT MAPPED IMAGES USING SALIENCY Kiran B. Raja Marius Pedersen The Norwegian Colour and Visual Computing Laboratory Gjøvik University College, Norway. kiran.raja@hig.no marius.pedersen@hig.no ABSTRACT Detecting artifacts introduced by gamut mapping algorithms is necessary to ensure the quality of color image reproduc- tion. Machine based detection of artifacts shall reduce the tedious work of visual inspection. In this work, we propose to use contrast information to detect the artifacts introduced due to the process of gamut mapping. We further evaluate the proposed algorithm on a set of gamut mapped images and an- alyze the results. The results are validated against the existing benchmarks. Index Terms— Gamut Mapping, Artifacts, Saliency, His- togram Contrast, Region Contrast 1. INTRODUCTION Reproducing images on different devices introduces artifacts into the images. Some of the well known artifacts can be exemplified as contours, halos, block creation and loss of details [1, 2, 3]. Detecting these artifacts in the images is necessary to ensure high quality. Many of the artifacts found in a workflow are due to limited number of colors a system can reproduce, in other terms, limited gamut availability. Existing work by Pedersen et al. [4] show that observers looked for artifacts in 40 percent of the images when judging the quality of printed images. Bonnier et al. [5] found that artifacts strongly influence the quality of gamut mapped im- ages. This work has also shown that detail is one of the most important quality attributes used when judging the quality of gamut mapped images. Hardeberg et al. [6] conducted an experiment to judge gamut mapping algorithms and the results showed that observers stated loss of shadow details as the most important criteria for the evaluation of gamut mapped images. Bara´ nczuk et al. [7] state that preserving spatial details, in addition to preserving color/lightness, is the main quality factor for gamut mapping. In experiments by Dugay et al. [8] results from expert observers showed that good rendering of details was an important criterion for obtaining high quality. Considering the amount of labor, resources and time involved in conducting the psychovisual experiments to evaluate the quality of images in terms of artifacts, it becomes essential to automate the process of artifact detection. This can be done through image quality metrics. Standard metrics like MSE and PSNR are not suitable for detecting these artifacts, since they are simple point-wise metrics and do not incorporate any models of the human visual system, more advanced methods are needed [3]. One of the methods specifically designed for evaluation of gamut mapped images was proposed by Cao et al. [9] which was based on ”difference of saliency”. Saliency of an image can be described as the perceptual quality which captures the attention of the observer [10]. The idea by Cao et al. [9] was that the visual attention is affected to a greater extent in images if certain details are lost or additional de- tails are introduced. The changes to an image that occur in a gamut mapping procedure may either introduce saliency or remove the saliency in corresponding regions of the same image which are both unwanted in the reproduction. Thus, any artifact becomes detectable if it introduces or removes the saliency. Hence, on a general note, the difference of salien- cies between the original image and gamut mapped image should give the locations of artifacts in the image. The starting point was the saliency detector proposed by Achanta et al. [11], which is primarily based on the Difference- of-Gaussian (DoG) filter with band-pass mechanism. The authors [11] transform the images into the CIEL*a*b* space to take the effect of lightness on saliency into account. The saliency detection method adopted efficiently outputs full res- olution saliency maps, establishing well-defined boundaries of salient objects and disregard high frequencies arising from texture, noise and blocking artifacts. In this work, we propose to extend the artifact detection model proposed by Cao et al. [9] using more robust saliency detection method proposed by Cheng et al. [12]. Their approach considers spatial infor- mation and its enhanced contrast corresponding to different regions to produce spatially coherent high quality saliency maps. They also propose pixel based contrast information to generate the saliency maps. These saliency maps with better accuracy, can be used to estimate the artifacts in a robust way. 2013 Colour and Visual Computing Symposium (CVCS) 978-1-4799-0609-3/13/$31.00 c 2013 IEEE