Redefining Binarization and the Visual Archetype

Although binarization is considered passe, it still remains a highly popular research topic. In this paper we propose a rethinking of what binarization is. We introduce the notion of the visual archetype as the ideal form of any one document. Binarization can be defined as the restoration of the visual archetype for a class of images. This definition broadens the scope of what binarization means but also suggests ground-truth should focus on the foreground. A. Defining Binarization Binarization has almost become a dirty word in some parts of the Document Image Analysis (DIA) community. In the authors opinion the reason for this is not that it has been solved, not that it is an ill posed problem [1] although it is, as much as the fact that the assumption of high quality binarization is practically infeasible in many real world cases. Binarization is also usually required by methods that use smearing, morphological operations etc. Furthermore in a recent experimental evaluation of the contribution methods have across a full DIA pipeline by Laminroy et al. [2], a binarization method proved to be the one with the highest positive contribution. Some attempts at a definition of binarization have been made in the literature. Kavalieratou describes it as a method that discriminates foreground from background, thus, removing any kind of noise that obstructs the legibility of the document image [3]. Stathis et al. defined it as automatically converting the document images in a bi-level form in such way that the foreground information is represented by black pixels and the background by white ones in [4] and Shafait et al. give a very similar definition [5]. Shafait et al. define binarization Ntirogiannis et al. define it as ”the process that segments the document image into text and background by removing any existing degradations” [?] . What is more indicative, is It is indicative that several high impact papers addressing binarization as their main topic, describe it as selecting a threshold [7], or even wisely avoid defining at all [8]. In [1] Lopresti and Nagy use binarization as an example of an ill-defined problem. In recent years the DIBCO [9] competitions have become the standard for benchmarking binarization methods. By defining the problem with respect to a given ground-truth the discussion of what is binarization has been bypassed. Smith et al. in [10] address the question of bias in ground-truthing and propose monitoring the effect ground-truth has over method development. What is not addressed with benchmarking on specific datasets, is the extent to which a method needs tuning.