Estimation of Illuminant Chromaticity Based on CCD Camera Response Distribution in a Real-World Image

This paper proposes a method for estimating the illuminant chromaticity using the distributions of the camera responses obtained by a CCD camera in a real-world scene. Illuminant estimation using a highlight method is based on the geometric relation between a body and its surface reflection. In general, the pixels in a highlight region are affected by an illuminant geometric difference, camera quantization errors, and the non-uniformity of the CCD sensor. As such, this leads to inaccurate results if an illuminant is estimated using the pixels of a CCD camera without any preprocessing. Accordingly, to solve this problem, the proposed method analyzes the distribution of the CCD camera responses and selects pixels using the Mahalanobis distance in highlight regions. The use of the Mahalanobis distance based on the camera responses enables the adaptive selection of valid pixels among the pixels distributed in the highlight regions. Lines are then determined based on the selected pixels with r-g chromaticity coordinates using a principal component analysis(PCA). Thereafter, the illuminant chromaticity is estimated based on the intersection points of the lines. Experimental results using the proposed method demonstrated a reduced estimation error compared with the conventional method. Illuminant Estimation Method Color constancy is the attempt to derive the intrinsic reflectance properties of objects, which are independent of extrinsic parameters, such as illumination, viewing, direction, surface orientation, and surrounding colors. In the case of human beings, the original color of an object under an arbitrary illuminant is estimated as an integrated judgment. However, an input device, such as a camera, is unable to discriminate the feature of an illuminant, as it only records the features of the original input responses. As such, there is a need for illuminant estimation to replicate the visual ability of humans. Illuminant estimation methods can be classified according to whether they use spectral reflectance or a tri-stimulus. The basic approach using spectral reflectance was originally designed by Maloney et. al. The spectral reflection from an object surface comes from the multiplication of the body and the surface reflection. Therefore, the spectral reflectance of an illuminant can be estimated using an analysis of this multiplication formula. D’Zmura et. al. proposed general linear and bilinear models to extend Maloney’s approach based on the combination of multiple illuminants and multiple surfaces and the relationship between these two factors. Plus, Tominaga proposed illumination estimation using singular value decomposition. In contrast, Land’s Retinex theory is the basic approach using a tri-stimulus input along with the grey world assumption that the average vector for the three channels is assumed to be the illuminant chromaticity for the scene or image. Other approaches using highlights have also been considered. Lee proposed a method to estimate the illuminant chromaticity by analyzing regions with a chromaticity change, i.e. for highlight regions in an image, the chromaticity distribution of the highlight region makes a line, and if there are more than two lines, the cross point is assumed to be the illuminant chromaticity. However, since conventional methods basically estimate an illuminant with either a synthetic or optimal image, it is difficult to obtain a good result for a real-world scene, as camera responses include quantization errors and non-uniform CCD sensors. To overcome this problem, Lehmann(color line search: CLS) recently proposed an illuminant estimation method for realworld scenes that uses additional captured images to compensate for camera noise. Yet, it is still difficult to apply this method to a real situation. Therefore, this paper proposes an illuminant estimation method using the Mahalanobis distance that considers the camera response distribution within a single image. Mahalanobis Distance Method The Mahalanobis distance indicates the relation between clusters or the relation between a cluster and a pixel, where d is the Mahalanobis distance between an arbitrary pixel and the centroid, S is the location vector of an arbitrary pixel, S is the mean vector of the training set, and Σ is the variance-covariance matrix for the training set. CGIV 2004: The Second European Conference on Colour Graphics, Imaging and Vision