IJDAR (2008) 11:39–51 DOI 10.1007/s10032-008-0068-2 ORIGINAL PAPER Robust reconstruction of low-resolution document images by exploiting repetitive character behaviour Hiêp Q. Luong · Wilfried Philips Received: 10 April 2008 / Revised: 27 July 2008 / Accepted: 4 August 2008 / Published online: 2 September 2008 © Springer-Verlag 2008 Abstract In this paper, we present a new approach for reconstructing low-resolution document images. Unlike other conventional reconstruction methods, the unknown pixel values are not estimated based on their local surroun- ding neighbourhood, but on the whole image. In particu- lar, we exploit the multiple occurrence of characters in the scanned document. In order to take advantage of this repe- titive behaviour, we divide the image into character seg- ments and match similar character segments to filter relevant information before the reconstruction. A great advantage of our proposed approach over conventional approaches is that we have more information at our disposal, which leads to a better reconstruction of the high-resolution (HR) image. Experimental results confirm the effectiveness of our propo- sed method, which is expressed in a better optical character recognition (OCR) accuracy and visual superiority to other traditional interpolation and restoration methods. Keywords Repetition · Restoration · Interpolation · Character segmentation · Bimodal distribution · OCR 1 Introduction Many applications nowadays rely on digital image interpola- tion. Especially improving image text resolution has become important in recognition applications. Some examples are improving the readability (e.g. of license plates provided by surveillance cameras or for office automation) and H. Q. Luong (B) · W. Philips Department of Telecommunications and Information Processing, IPI, IBBT, Ghent University, Sint-Pietersnieuwstraat 41, 9000 Ghent, Belgium e-mail: hiep.luong@telin.ugent.be simple spatial magnification (e.g. printing low-resolution documents on high-resolution printer devices or displaying text in low-resolution pictures on the next-generation e-papers or High Definition televisions (HDTV) screens). Optical character recognition (OCR) is a useful tool in digi- talizing libraries, computer-assisted indexing and retrieval of video archives, etc. However text observed in low resolution (e.g. in poor quality video or with very small fontsize) reduces the OCR performance heavily. That is why we need docu- ment reconstruction methods in order to improve the OCR accuracy. Many image interpolation methods, which are not desi- gned specifically for text, have already been proposed in the literature, but all suffer from one or more artefacts especially when applied to text. Linear or non-adaptive interpolation methods cause staircasing (i.e. jagged edges in the upscaling process), blurring and/or ringing effects. Well-known and popular linear interpolation methods are the nearest neigh- bour method, bilinear interpolation, interpolation with higher order (piecewise) polynomials, B-splines, truncated or win- dowed sinc functions, etc. [20, 28]. Adaptive interpolation methods incorporate prior know- ledge about images. Some methods focus on reconstructing edges [22, 30] and other methods tackle unwanted interpola- tion artefacts such as staircasing, blurring and ringing using isophote smoothing, level curve mapping or mathematical morphology [18, 25, 29]. Another class of adaptive image enlargement methods is the training-based approach, which maps blocks of the low-resolution image into predefined high-resolution blocks [14]. This has been successfully applied to text images [8, 9]. However, the results depend heavily on the used training set and thus the font type (which must be known in advance). Other specific text enhance- ment methods focus on contrast improvement [7], pixel patterns [41], fixing broken or touching characters [1, 4], 123