Document sorting and logo recognition using image processing techniques
Scott Lowther · QUT ePrints (Queensland University of Technology) · 2002
In recent years, document analysis has become an ever-increasing area of image processing. This has been primarily due to the interest in the notion of a paperless office but also due to the exponential increase in personal computer power. Document analysis broadly covers such areas as document pre-processing, Optical Character Recognition (OCR), script/language recognition and document understanding. One important goal of many document analysis systems is to automatically process document images and to extract meaningful information. Systems such as document sorting applications aim to automatically sort large volumes of paper (or scanned) documents into smaller groups to allow human operators to quickly index document databases for relevant documents. The development of such a system relies on selection of numerous other document analysis tasks such as pre-processing, script I language recognition, OCR, image recognition and document layout analysis, dependent upon the information of interest. This thesis presents a study of document analysis techniques applicable to a document sorting application using logo recognition. The primary goal of the research performed was to develop algorithms that can recognise logo images in documents and sort these documents according to corporate identity. The thesis outlines many of the current techniques required for sorting documents by corporate identity and presents some original techniques that were developed to address problems with current techniques. The areas investigated within the scope of this thesis are document pre-processing, logo detection and logo recognition, all of which are used extensively in a logo recognition document sorting system. An analysis of the pre-processing tasks of noise removal, binarization and skew determination is made and a new skew determination technique presented. A rule-based logo detection technique is developed and results provided. Bispectral invariant features are analysed for suitability in logo recognition and a technique developed to both recognise and verify logo images. The integration of all these techniques is then examined and a number of compatible techniques chosen for inclusion in a complete system. Results are provided to show the validity of each of the main techniques used and also of the complete system.