A novel method using "color drop-out" for document images with "color shift" is proposed. Color shift phenomena sometimes occur in document images captured by a camera device or stand type scanner. It adversely affects the binarization and character recognition processes, because it generates pseudo color pixels on scanned image, which do not exist on the original document. To solve the "pseudo color problem," a binarization method based on the following three calculation steps is proposed. First, line and character areas are estimated coarsely by using form structure analysis and subtracting background from images, second, the color shift is removed by using morphological processing, third, each pixel of the background subtracted images is discriminated into character strings and lines precisely by dynamic color classification. Several character recognition experiments using low quality form samples (in which handwritten strings overlap form lines and preprints) were performed. According to the experimental results, the proposed method attains character recognition accuracy of 94.3%, which is 5.2pt. higher than that attained by a conventional method.