Copyrightc 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR).Tone Reproduction: A Perspective from Luminance-Driven Perceptual GroupingHwann-Tzong Chen Tyng-Luh Liu Tien-Lung ChangInstitute of Information Science, Academia SinicaNankang, Taipei 115, Taiwan{pras, liutyng, tino}@iis.sinica.edu.twAbstractWe address the tone reproduction problem by integratinglocal adaptation with global-contrast consistency. Manyprevious works have tried to compress high-dynamic-range(HDR) luminances into a displayable range in imitation ofthe local adaptation mechanism of human eyes. Neverthe-less, while the realization of local adaptation is not theo-retically defined, exaggerating such effects often causes un-natural global contrasts. We propose a luminance-drivenperceptual grouping process to derive a sparse representa-tion of HDR luminances, and use the grouped regions toapproximate local properties of luminances. The advantageof incorporating a sparse representation is twofold: We cansimulate local adaptation based on region information, andsubsequently apply piecewise tone mappings to monotonizethe relative brightness over only a few perceptually signif-icant regions. Our experimental results show that the pro-posed framework gives a good balance in preserving localdetails and maintaining global contrasts of HDR scenes.1. IntroductionThe attempt to reproduce the visual perception of the realworld is at the heart of painting and photography. Artistshave long been endeavoring to develop skills in simulatingactual reflected light within the limitation of the medium,since our world generally delivers a much wider range ofluminances than pigments can reflect. Apart from artisticconcern, recreating real-scene impressions on limited mediais also inevitable in many vision and graphics applications.For example, the highest contrast of today’s LCD monitorsis around 1, 000 to 1; however, we may still need to displaya sunset scene whose contrast exceeds 10, 000 to 1.By a tone reproduction problem, we focus on establish-ing an efficient method to faithfully reconstruct the high-dynamic-range (HDR) radiance on a low-dynamic-range(LDR) image. The dynamic range of a digital image is sim-ply the contrast ratio in intensity between its brightest anddarkest parts. In [20], Ward introduces a floating-point pic-ture format to record HDR radiance in 32 bits per pixel, anddesigns a graphics rendering system that outputs images inthat format. Debevec and Malik have shown that HDR ra-diance of real scenes may be captured using regular SLRor digital cameras [5]. They propose a method to combinea series of pictures with different exposure settings into asingle HDR image, which is called a radiance map, withcontrast of about 250, 000 to 1. In this context, the aim ofour research can be stated as solving the tone reproductionproblem of radiance maps, that is, generating a displayablestandard RGB image that preserves perceptual properties ofthe original HDR radiance map.1.1. Related WorkSeveral works have been devoted to producing HDR imagesof real scenes [2], [5], [11], [12], including those that aredesigned to capture HDR luminances simultaneously undermultiple exposures[2], [11]. Inherently, panoramic imagingcan also be extended to carry HDR luminances by addingspatially varying optical filters to a camera [1], [16]. Dif-ferent from static HDR imaging, Kang et al. [9] proposeto generate HDR videos by changing the exposure of eachframe and then by stitching consecutive frames.On displaying HDR images, tone reproductionaddressesvisibility and impression through finding an appropriatemapping to compress high contrasts into a visible (dis-playable) range, accounting for perceptual fidelity. With aglobal mapping, pixels are mapped uniformly regardless oftheir spatial or local properties, and hence details are oftensmeared. The main advantage of using a global mappingis its efficiency. Ward et al. [21] describe a more sophis-ticated method to adjust contrast globally based on lumi-nance histograms. Nevertheless, the approach still smoothsout the details in areas of flat histograms. To improve visualfidelity, a number of tone reproduction methods have ex-plored nonuniform (local) mappings, as human visual sys-tem operates more likely this way [3], [6], [7], [8], [14],[19]. In particular, visual cells are organized in a center-surround manner so that we can see a wide range of lumi-nances by discriminating locally. In simulating the center-surround organization, Reinhard et al. [14] calculate, foreach pixel, the average intensity of a proper circular region,and then use the information to adjust a mapping function.Decomposing a radiance map into layers is another pop-ular choice for preserving image details. Methods of thiskind often separate an image into an illumination layer anda reflectance layer. The illumination layer carries the lu-minance information of the original image and thus hasa wider dynamic range, while the reflectance layer keepsthe textures and is of low dynamic range. Consequently,the dynamic range of an HDR image can be reduced bycompressing its illumination layer. For images of naturalscenes, Land’s retinex theory [10] can be used to estimateillumination and reflectance. Indeed center-surround basedand layer-decomposition based methods are closely related.Both aim to preserve details and exploit local adaptation tomatch human perception. However, overemphasizing localcontrasts may produce halos, which are defects of reversalcontrasts. A number of new methods have been introducedto resolve halos by incorporating more appropriate local av-eraging schemes, e.g., bilateral filtering [18] used in [6],[7], multi-scale Gaussians [3], and dodging-and-burning[14], or, by directly working on the gradient domain accord-ing to derived PDE formulations, e.g. anisotropic diffusion[19] and the Poisson equation [8].From a segmentation viewpoint, Schlick [17] has pro-posed to divide an image into zones of similar values, andthen compute the average intensity of each zone. The av-erage intensity map can be used to constitute the spatiallynonuniform tone mapping function. Yee and Pattanaik [22]develop a multi-layer partitioning and grouping algorithmto compute the local adaptation luminance. In each layer,pixels are partitioned based on a given intensity resolution(bin-width), and pixels that are partitioned into the same binform a group. Each pixel’s local adaptation