Scanned color document image segmentation using the EM algorithm

A robust, efficient scanned color document segmentation algorithm is presented that performs a three-dimensional (31)) thresholding of color pixels. At the heart of the algorithm is the Expectation-Maximization (EM) algorithm which fits a mixture of two 3D gaussians to L* a* b* color data sampled from pixels in the image to separate foreground and background. The thresholding process uses a quadratic boundary to produce a selector plane that indicates, for each pixel, whether it belongs to foreground or background. The binary selector plane is further processed and analyzed to extract objects such as photographs or graphics. Segmented document images are then encoded in a mixed raster content format for efficient compression.