Robust Color Image Enhancement of Digitized Books

We present a spatially variant framework for correcting uneven illumination and color cast, problems commonly associated with digitized books. The core of our method is a color image segmentation algorithm based on watershed transform and noise thresholding. We propose a threshold estimation method using the histogram of the gradient magnitude. The critical parameters of the segmentation are further constrained by statistical analysis of multiple book pages conducted in a bootstrap pass. We demonstrate the consistent performance of the proposed method on books with wide-ranging content and image quality.