Binarization of Camera-Captured Document using A MAP Approach Xujun Peng, Srirangaraj Setlur, Venu Govindaraju and Ramachandrula Sitaram † CUBS, University at Buffalo, Amherst, NY 14228, USA † HP Labs India, Bangalore 560030, India ABSTRACT Document binarization is one of the initial and critical steps for many document analysis systems. Nowadays, with the success and popularity of hand-held devices, large efforts are motivated to convert documents into digital format by using hand-held cameras. In this paper, we propose a Bayesian based maximum a posteriori (MAP) estimation algorithm to binarize the camera-captured document images. A novel adaptive segmentation surface estimation and normalization method is proposed as the preprocessing step in our work and followed by a Markov Random Field based refine procedure to remove noises and smooth binarized result. Experimental results show that our method has better performance than other algorithms on bad or uneven illumination document images. Keywords: Document Binarization, Markov Random Field, Image Processing, Camera-captured 1. INTRODUCTION Because most document retrieval and recognition methods rely on high quality binarized document images, the document image binarization which segments foreground text area from blank background plays a key role in document analysis and recognition system. With the vast amount of using of the hand-held devices, more and more documents can be captured and converted into digital format using hand-held devices such as PDA and hand-held camera with high-resolution. However, despite the convenience of the hand-held devices, the camera- captured document images easily suffer the degradation such as distortion, uneven or bad illumination, etc which cause the binarization be a challenge problem. The research into the image binarization can be traced back to the pioneering work on global threshold based methods, such as the famous Otsu’s method 1 which used a single threshold value that maximizes the inter-class (foreground and background) variance or minimizes the intra-class variance to segment the entire image. The drawback of this method is that it assumes the histogram of images has two distinct peaks for different classes respectively and can be separated. But this assumption is hardly satisfied in most real applications, especially to camera-captured document images. Fig. 1 shows examples of degraded camera-captured document images, along with their corresponding histograms. To overcome the disadvantage of single global threshold binarization method, Valizadeh et al. 2 suggested mapping the original grey level of each pixel to a new domain prior to using global threshold. Shi and Govin- daraju 3 normalized grey level of pixels within degraded historical document image and binarized them through the global thresholding. By using a two rounds polynomial surface smoothing process, Lu and Tan 4 used similar manner of normalization before thresholding. Another thrust has been on using the adaptive threshold or local threshold which computes the threshold of each pixel according to the properties of its own or its neighbor pixels. Niblack 5 proposed a method which uses mean value and variance of small window to determine the threshold of centered pixels. The potential problem of this method is that large amount of noises are produced in pure blank background areas and it is sensitive to the window size. An extension of Niblack’s method was described by Sauvola et al. 6 which obtained better performance in open background regions. In Gatos’s 7 work, a foreground and background surface estimation algorithm was proposed followed by an adaptive thresholding procedure. Instead of using Otsu threshold globally, E-mail: {xpeng,setlur,govind}@buffalo.edu and sitaram@hp.com Document Recognition and Retrieval XVIII, edited by Gady Agam, Christian Viard-Gaudin, Proc. of SPIE-IS&T Electronic Imaging, Vol. 7874, 78740R · © 2011 SPIE-IS&T CCC code: 0277-786X/11/$18 · doi: 10.1117/12.874091 SPIE-IS&T/ Vol. 7874 78740R-1 Downloaded from SPIE Digital Library on 29 Jan 2011 to 128.125.140.80. Terms of Use: http://spiedl.org/terms