Fig 5 - uploaded by Costin-Anton Boiangiu
Content may be subject to copyright.
The difference between a normal letter (a) and a bold letter (b) regarding the ratio between the number of outline pixels and the total number of pixels.
Source publication
This paper describes an approach towards obtaining a normalized measure of text resemblance in scanned images, relying on the detection of standard character features, and using a sequence of procedures and algorithms on input images, for automatic content conversion purposes. The approach relies solely on geometrical characteristics of the charact...
Similar publications
Self-similarity based super-resolution (SR) algorithms are able to produce visually pleasing results without extensive training on external databases. Such algorithms exploit the statistical prior that patches in a natural image tend to recur within and across scales of the same image. However, the internal dictionary obtained from the given image...
Citations
... Boiangiu et al. [6] proposed another method for the classification of text based on geometrical characteristics. The method was tested on roman script newspapers. ...
... The components may be represented by their bounding boxes or the pixels themselves or by Voronoi diagrams, Delaunay triangulation or 3D meshes. [70][71] [72] PRIMITIVE LEVEL PROCESSING The aim of primitive level processing is to identify the nature of the elements that have become accessible following the preprocessing steps. Whereas for text documents the focus is the eventual recognition of characters, for non-textual documents the aim is recognizing basic geometric entities. ...
... Grouping the different elements is a matter of observing distance conventions that are usually respected while writing texts, finding interpreting white spaces, lines [75] and any form separators [76]. Methods for detecting font characteristics accurately give valuable information [70]. Authors evaluate these distances relative to text size [77], as the blank columns separating two text columns need to be large enough to unambiguously split two lines that have the same vertical coordinates [42]. ...
This paper presents an overview of Document Image Analysis Systems, their composing modules, the approaches these modules use, as well as uses for these applications. One of the main goals is to present some of the most important technologies and methods behind the Document Image Analysis domain in order to evaluate the best approach when dealing with real-world documents. The other main goal is to ensure a foundation for those starting to build such complex software systems and to give an elaborate technical answer to the question: “How to make physical documents available to a large number of people?”
... The components may be represented by their bounding boxes or the pixels themselves or by Voronoi diagrams, Delaunay triangulation or 3D meshes. [70][71] [72] ...
... Grouping the different elements is a matter of observing distance conventions that are usually respected while writing texts, finding interpreting white spaces, lines [75] and any form separators [76]. Methods for detecting font characteristics accurately give valuable information [70]. Authors evaluate these distances relative to text size [77], as the blank columns separating two text columns need to be large enough to unambiguously split two lines that have the same vertical coordinates [42]. ...
... The components may be represented by their bounding boxes or the pixels themselves or by Voronoi diagrams, Delaunay triangulation or 3D meshes. [70][71] [72] ...
... Grouping the different elements is a matter of observing distance conventions that are usually respected while writing texts, finding interpreting white spaces, lines [75] and any form separators [76]. Methods for detecting font characteristics accurately give valuable information [70]. Authors evaluate these distances relative to text size [77], as the blank columns separating two text columns need to be large enough to unambiguously split two lines that have the same vertical coordinates [42]. ...
... The construction of independent modules of analyze and image processing (image digitization, image segmentation, skew detection and correction, layout analyze) Interfacing the various modules to generate a full processing pipeline (edit interconnectivity graph, run multiple approaches for the same recognition problem and use voting to validate and improve the results) Editing of results for the various intermediate steps in layout and hierarchy detection, OCR; thus allowing human operators to verify and correct (if necessary) the output of different heuristically-based algorithms Import and export of document data In order to effectively use the obtained results the files will be generated in PFD format, structured in a multi-layer manner: the original image stays in the front layer (to ensure that the original of the document is preserved and seen by the user) and the meta-information in a background layer (to ensure the usability of document in search, classification and textretrieval operations). Input preparation -scan/import of documents Image Preprocessing -skew detection and compensation, digitization and/or image segmentation (Srinivasan and Shobha, 2008) Layout Analyze -clustering of the image segments resulted in previous steps, according to the logical reading order of the document (Boiangiu, 2008) Page Numbering -markers detection and renumbering, if necessary Hierarchy Analyze -layout blocks are classified according to their importance and the logical table of content for the document is constructed Optical Text Recognition -text areas are processed to extract the valid words/sentences and their formatting Output/Export of electronic document in various formats Furthermore, the proposed service wants to combine the best of research and industry worlds: assuring both the possibility to add components and user interfaces, specific to a commercial service which also considers usability and a research oriented open source product in a fully functional service. ...
The main goal of this paper is to present the architecture and functionality of e-Service based platform. The project is structured along several dimensions that follow the development of complementary services, integrated to support everyday work experience, research and learning in the University POLITEHNICA of Bucharest (UPB). The platform support is represented by Internet as a large scale distributed environment. The current evolution of the Internet can be viewed from multiple perspectives: service oriented (Internet of Services), users centered (Internet of People), real-world integration over Internet (Internet of Things), production and use of multimedia content over Internet. The main services in eUPB platform are: (i) data retrieval, aggregation and search service, (ii) communication service for heterogeneous networks, (iii) mobile services to support context-aware applications, (iv) secure data delivery infrastructure for wireless sensor networks, (v) 3DUPB - the 3D MMO virtual replica of UPB, (vi) analysis and content extraction of scanned documents, and (vii) collaboration service. This is a position paper presenting the general architecture of eUPB and a description of each services device and functionality.
This paper describes a collection of algorithms for detecting text areas in document images using morphological operators, text clustering using geometrical text measurements and efficient image coding for fast remote correction in automatic content conversion systems Text characteristics are automatically discovered and used to filter out all non-text areas in the image. All the algorithms were implemented and tested on a representative set of test images obtained by scanning newspapers, books and magazines. The document image page clustering uses a measure of normalized text font resemblance. The approach makes use solely of the geometrical characteristics of characters, ignoring information regarding context or character recognition.
Entity clustering is a vital feature of any automatic content conversion system. Such systems generate digital documents from hard copies of newspapers, books, etc. At application level, the system processes an image (usually in black and white color mode) and identifies the various content layout elements, such as paragraphs, tables, images, columns, etc. Here is where the entity clustering mechanism comes into play. Its role is to group atomic entities (characters, points, lines) into layout elements. To achieve this, the system takes on different approaches which rely on the geometrical properties of the enclosed items: their relative position, size, boundaries and alignment. This paper describes such an approach based on 3D mesh reduction.
In an image processing application there is often need to identify and extract different morphological elements such as characters or lines. This paper studies one general method of identifying vertical or horizontal lines. However, the techniques described here can be used to detect other analytical objects like circles, or even ellipses. Line Detection is an important add-on to an automatic content conversion system which builds digital documents from scanned papers. After identifying lines, other layout elements can be extracted: columns, paragraphs, tables, headers. The present paper is a study of the Hough Transform for which several new enhancements are introduced.
This paper describes an approach towards obtaining the normalized measure of text resemblance in scanned images. The technique, aimed at automatic content conversion, is relying on the detection of standard character features and uses a sequence of procedures and algorithms applied sequentially on the input document. The approach makes use solely of the geometrical characteristics of characters, ignoring information regarding context or the character-recognition.
