Segmentation of Postal Envelopes for Address Block Location: an approach based on feature selection in wavelet space David Menoti, Díbio Leandro Borges, Jacques Facon, Alceu de Souza Britto Jr Pontifical Catholic University of Parana (PUCPR) Postgraduate Program in Applied Informatics (PPGIA) Image Science Group Rua Imaculada Conceição, 1155 – Prado Velho, Curitiba, PR, Brazil E-mail: { menoti, dibio, facon, alceu } @ppgia.pucpr.br Abstract This paper presents a segmentation algorithm based on feature selection in wavelet space. The aim is to automatically separate in postal envelopes the regions related to background, stamps, rubber stamps, and the address blocks. First, a typical image of a postal envelope is decomposed using Mallat algorithm and Haar basis. High frequency channel outputs are analyzed to locate salient points in order to separate the background. A statistical hypothesis test is taken to decide upon more consistent regions in order to clean out some noise left. The selected points are projected back to the original gray level image, where the evidence from the wavelet space is used to start a growing process to include the pixels more likely to belong to the regions of stamps, rubber stamps, and written area. Experiments are run using original postal envelopes from the Brazilian Post Office Agency, and here we report results on 440 images with many different layouts and backgrounds. 1. Introduction Postal automation has been recently integrated into the research agenda of the pattern recognition and computer vision communities, since acquisition and storage of images of envelopes and parcels has become easier and cheaper than a decade ago. However, segmentation of a typical image of a mail piece into background, stamps, and the address blocks is still a challenging problem due also to the large variety of stamps, backgrounds, written text of the address (e.g. handwritten, printed, locations). Other works in the literature have tackled different aspects of that problem. A survey in document image understanding up to 1994 can be seen in [3]. In [1] a texture segmentation technique, which organizes the wavelet coefficients of an image into a probabilistic graph is presented. Fusion of that information by Hidden Markov modeling is used to refine segmentation hypotheses. A layout page segmentation is presented in [2], and it is based on local feature extraction by wavelet packets, followed by a soft integration process to vote for layout borders detection. One of the few works we found with results on envelopes is in [4], which presents a method to identify regions in envelope images candidates for being the destination address. The technique is a texture segmentation based on Gabor filters. In [5] a method to locate text areas against different backgrounds is shown, which is based on a pseudo-motion technique to identify oscillations on the wavelet coefficients. An integrated system which performs fast identification of zip codes given that the address block is provided is reported in [7]. A parsing algorithm works intensively separating words and symbols represented contours in the form of chain codes. In [9] a fast segmentation approach for address block location on oversized flat envelopes was presented. The approach works based on measuring homogeneities of gray level blocks, and adaptive threshold values from the gradients of the blocks. An split and merge like algorithm is then presented in [10] which works based on special geometric layouts for the address block. Corners and orientation of the lines are extracted and that information is used for separating background features from the address block. An interesting work in text detection in document images such as newspapers, photographs, and magazines is shown in [11], where a texture segmentation module uses gaussian derivative filters followed by a non-linear transformation to produce the feature vectors. A method to locate address blocks on images where an arbitrary layout of printed text is known a priori is presented in [12]. We present here a novel approach for segmentation of an image of a postal envelope, it is a general and robust segmentation method not restricted to a particular layout. Proceedings of the Seventh International Conference on Document Analysis and Recognition (ICDAR 2003) 0-7695-1960-1/03 $17.00 © 2003 IEEE