A method for binarizing a scanned document images containing gray or light colored text printed with halftone patterns. The document image is initially binarized and connected image components are extracted from the initial binary image as text characters. Each text character is classified as either a halftone text character or a non-halftone text character based on an analysis of its topology features. The topology features may be the Euler number of the text character; a text character with a Euler number below −2 is classified as halftone text. The gray-scale document image is then divided into halftone text regions containing only halftone text characters and non-halftone text regions. Each region is binarized using its own pixel value statistics. This eliminates the influence of black text on the threshold values for binarizing halftone text. The binary maps of the regions are combined to generate the final binary map.