I N D O C U M E N T I M a G E S

A method for analyz ing the structu re of the white backgr ound in docu ment images is describ ed, along with applic ations to the proble m of isolati ng blocks of machine -printe d text. The appro ach is based on computa tionalgeometry algorith ms for off-li ne enumera tion of maximal white rectan gles and on-line rectan gle unific ation. These suppo rt a fast, simple, and gener al heuri stic for geometr ic layout segment ation, in which white space is covere d greed ily by rectan gles until all text blocks are isolated . Desig n of the heuri stic can be substa ntially automate d by an analysi s of the empiric al statisti cal distribu tion of propert ies of coveri ng rectan gles: for example , the stoppin g rule can be chosen by Rosenblatt ’s percep tron traini ng algorith m. Experi mental trials show good behavio r on the large and useful class of textua l Manhatt an layouts . On complex layouts from Englishlangua ge technic al journa ls of many publis hers, the method finds good segmentations in a unifor m and nearl y parame ter-fre e manner . On a variety of non-La tin texts, some with vertic al text lines, the method finds good segmenta tions witho ut prior know ledge of page and text-l ine orienta tion.