Character Degradation Model and HMM Word Recognition System for Text Extracted from Maps
Aria X. Pezeshk, L. Richard · InTech eBooks · 2011
Will-be-set-by-IN-TECH sections of the map and editing capabilities are supported in this case, heads-up digitizing is now the more popular method for digitizing the graphical features in paper maps.In addition to the graphical elements, the textual content of maps also needs to be manually extracted and assigned to the appropriate objects in the digital copy.Manual conversion of maps into usable digital format is therefore a tedious, labor intensive, and error prone process.While these procedures may be viable for small scale operations that only involve a few maps, they are unpractical for large scale conversion of whole map repositories.A system that can automate this process is therefore essential for GIS users to replace the manual map conversion methods and provide them with easy access to the vast number of printed maps that exist in various archives and produced by different organizations around the globe.In this chapter we discuss a custom recognition system for text extracted from maps.This work is part of a larger system that enables automatic sequential extraction of various graphical features, and subsequently processes the remaining image for text grouping, reorientation, and extraction (Pezeshk & Tutwiler, 2008a;b;2010b).The recognition engine presented here extends the application of the system we developed in (Pezeshk & Tutwiler, 2010a) by enabling multi-font and segmentation-free recognition.We focus on this particular problem because unlike other parts of the map conversion process, text recognition has received little attention in the research carried out in the field of automatic map understanding systems (as will be discussed in Section 3).Moreover, the level of noise and deformities in text extracted from maps can be too severe for commercial Optical Character Recognition (OCR) systems, resulting in low recognition rates and thereby higher user involvement in correcting the errors in post processing.The system proposed here can find a variety of applications.Combined with our previous work in (Pezeshk & Tutwiler, 2008a;b;2010b), we obtain a complete map understanding system that can digitize both the graphical and textual content of printed maps.Since both types of features are extracted at the same time, the text recognized using our system can easily be associated with the other digitized objects in the image (e.g.road labels with the corresponding road segments) in order to obtain an exact digital representation.Moreover, the recognized text can be entered into a database and used to search for maps that contain a particular street or area of interest, or to classify map archives in a similar manner according to their text content.Since geographic maps are accurately geo-referenced, the text can also be used in conflation with satellite or aerial imagery to automatically tag locations or objects that exist in these types of images.The remainder of this chapter is organized as follows.First, in Section 2 we discuss the challenges encountered in the extraction and recognition of the text content of scanned maps.A brief overview of research in the fields of document image analysis and map conversion algorithms is then presented in Section 3. Details of the different components of our text recognition system which include a custom noise model for the generation of artificial training sets, preprocessing algorithm for automatic text size normalization and orientation correction, and the Hidden Markov Model (HMM) based text recognition engine are subsequently discussed in Sections 4.2, 4.3, and 4.4, respectively.Finally, experimental results are shown in Section 5 before concluding the chapter in Section 6. ChallengesGeographic maps are designed to convey the largest possible amount of information and details regarding both the natural features (e.g.rivers and vegetation cover) and manmade 54 Recent Advances in Document Recognition and Understanding www.intechopen.com55