Image Quality and Change of Illuminant: An Information-Theoretic Evaluation

摘要: We investigate the influence of scene illuminant on perceived image quality. Given two multispectral images, an original and a reproduction (e.g. compressed, gamut mapped,..), we seek redundancies difference through changes illuminant, with regard to 5 socalled features (IDF). In order do this, employ information-theoretic perspective measure variations entropies in each IDF, w.r.t. various illuminants, case particular kinds distortions: spectral mapping reconstruction from six-channel camera model. Our results indicate that changing has lesser achromatic features. Introduction Most recent studies Image Quality Assessment (IQA) rely greyscale [1, 2] or chromatic information [3] rate between images. The intent is correlate as much possible human’s judgment under specific viewing conditions. Yet, advent technologies, appearance models multichannel printing, there growing need for higher dimensional IQA. Spectral acquisition, processing methods [4, 5], separation [6], compression [7], dimensionality reduction [8, 9]), require new range measures (SIQA). Although criteria such classification target detection accuracy are widely used quality remote sensing applications, very little work actually been done evaluate terms perception. [10], [11] [12], spectral-based distances divergence were studied hyperand quality, attempts relate these quantities perceptual meanings. they might human some cases, operate directly space popular RootMean Square Error Goodness-of-Fit Coefficient, usually unable properly so. A reasonable explanation this notion color (at least perception) exists only when Viewing Conditions (VC) specified. Without considering at observer model, no assumptions can be made how reflectance spectra interpreted by visual system. An alternative strategy pool scores traditional like CIE2000 over variety VC [13], but which extent? How much, aspects images remain unchanged one set another daylight incandescent light)? study, questions, better understand key challenges SIQA. propose use image-difference introduced Lissner et al. [3], observe their behavior changes. [2, 14] feature, Note most influential feature certainly power distribution illuminant. Therefore, study focuses solely while remaining standard observer) assumed constant. differences changed Image-Difference Features (IDF) Color ImageDifference (CID) [3]. CID framework, compare first normalized imageappearance including CAT02 adaptation (as CIECAM02). This take into account “the system’s capability adjust varying colors illumination approximately preserve object colors” [16]. then converted nearly perceptually uniform LAB2000HL [17]. IDF maps computed means adapted SSIM index [1] within sliding windows (see formulas Appendix). Five therefore obtained: Lightness-Difference map (LL), Lightness-Contrast (CL), Lightness-Structure (SL), Chroma-Difference (LC) Hue-Difference (LH). Figure 1 illustrates workflow extraction. refer paper source code provided [18] further explanations about measure. compares tri-chromatic intrinsically linked certain conversion data CIEXYZ tristimuli, was CIE 2◦ observer, illuminants. previous, [15], observed few illuminants able predict large variety. aim