Cast Shadow Identification and Image Restoration by Clustering Technique

In this paper, we present a novel approach for identifying and removing cast shadow in a color image. The technique employs clustering and color normalization procedures without the usual assumption that the darkest region constitutes a shadow or requiring the camera to be linear. The input image is transformed to a feature space spanned by the R-,G-,Bcolors and the Mean Shift Algorithm used for clustering. The number of such clusters denotes the number of significant distinct color regions including the shadow in the image. Color normalization has the tendency to remove shadow if a linear color space exists but does not if the space is non-linear. Using normalized color and Euclidean distance measure constraint, pairs of closest clusters called shadow candidate pairs are formed. Any shadow candidate pair whose distance apart is greater than the constraint is discarded as an invalid pair and vice-versa. The darker of the valid pair is regarded as shadow and can then be extracted. Our technique is also able to recover the background that is partly in shadow and partly illuminated. Image restoration is done by a mapping process whereby pixels that are resident in the shadow cluster are mapped to the mean of the ones they are closest to. We present some results using real color images with shadows on singleand multi-color backgrounds captured with a Kodak Zoom 120DC camera.