Stage-oe-small.jpg

Thema4675: Unterschied zwischen den Versionen

Aus Aifbportal
Wechseln zu:Navigation, Suche
 
Zeile 9: Zeile 9:
 
|Abschlussarbeitsstatus=Vergeben
 
|Abschlussarbeitsstatus=Vergeben
 
|Beginn=2021/04/01
 
|Beginn=2021/04/01
|Ausschreibung=Handwritten and Printed Text Separation in Historical Documents.pdf
+
|Ausschreibung=Handwritten and Printed Text Separation in Historical Documents_New Version.pdf
 
|Beschreibung DE=Objective of this work:  
 
|Beschreibung DE=Objective of this work:  
  
With the increase of digitized documents, automatic document analysis has become extremely important. The presentation of administrative documents for public introduces varieties of document types, content, quality and structure. Fundamentally speaking, documents can be skewed, noisy, overlapped with graphics, i.e., lines, unconstrained annotations, stamps.  
+
With the increase of digitized documents, automatic document analysis has become extremely important. The presentation of historical documents to the public introduces a variety of document types, content, quality and structure. Fundamentally speaking, documents can be skewed, noisy, and overlapped with graphics, i.e., lines, unconstrained annotations, stamps.  
In this thesis, existing technologies for visual semantic analysis of unstructured data will be investigated. The dataset of the work consists of 5595 images of tax forms from 1988. In order to provide a textual representation of the documents automatic text and pattern recognition processes have to be applied. An optical character recognition (OCR) system recognizes either printed or handwritten text. Hence, the task of the thesis is to separate machine printed text from handwritten text in scanned documents before feeding it to an OCR system. As a first step, the images will be preprocessed (e.g. cropping, noise filtering). Then the images will be segmented (e.g. line/word segmentation). As the last contribution the segments will be fed into a deep learning binary classifier to be separated into handwritten and printed text segments.  
+
Most optical character recognition (OCR) systems recognize either printed or handwritten text. Hence, the task of the thesis is to separate machine printed text from handwritten text in scanned documents before feeding it to an OCR system.
  
  
The project work will be supervised by Prof. Dr. Harald Sack, Tabea Tietz and Oleksandra Vsesviatska, Information Service Engineering at Institute AIFB, KIT, in collaboration with FIZ Karlsruhe.
+
In this thesis:
 +
 
 +
*Documents containing a mix of handwritten and printed  text will be collected.
 +
*An additional mixed dataset may be generated from historical documents.
 +
*The existing approaches of text separation will be reviewed and investigated.
 +
*A pixel-based approach for text separation based on [1] will be applied.
 +
*The results will be evaluated based on the ground truth data.
 +
 
 +
 
 +
[1] Dutly, N., Slimane, F., & Ingold, R. (2019, September). Phti-ws: A printed and handwritten text identification web service based on fcn and crf post-processing. In 2019 International Conference on Document Analysis and Recognition Workshops (ICDARW) (Vol. 2, pp. 20-25). IEEE.
 +
 
 +
 
 +
The project work will be supervised by Prof. Dr. Harald Sack, Mahsa Vafaie and Oleksandra Bruns, Information Service Engineering at Institute AIFB, KIT, in collaboration with FIZ Karlsruhe.
  
  
 
Keywords:  
 
Keywords:  
  
Knowledge Graphs, Cultural Heritage, NLP
+
Machine Learning, CNN, pattern recognition
 +
 
  
 
Pre-requisites:  
 
Pre-requisites:  

Aktuelle Version vom 25. März 2021, 15:30 Uhr



Handwritten and Printed Text Separation in Historical Documents




Informationen zur Arbeit

Abschlussarbeitstyp: Bachelor
Betreuer: Harald SackOleksandra VsesviatskaMahsa Vafaie
Forschungsgruppe: Information Service Engineering
Partner: FIZ Karlsruhe
Archivierungsnummer: 4675
Abschlussarbeitsstatus: Vergeben
Beginn: 01. April 2021
Abgabe: unbekannt

Weitere Informationen

Objective of this work:

With the increase of digitized documents, automatic document analysis has become extremely important. The presentation of historical documents to the public introduces a variety of document types, content, quality and structure. Fundamentally speaking, documents can be skewed, noisy, and overlapped with graphics, i.e., lines, unconstrained annotations, stamps. Most optical character recognition (OCR) systems recognize either printed or handwritten text. Hence, the task of the thesis is to separate machine printed text from handwritten text in scanned documents before feeding it to an OCR system.


In this thesis:

  • Documents containing a mix of handwritten and printed text will be collected.
  • An additional mixed dataset may be generated from historical documents.
  • The existing approaches of text separation will be reviewed and investigated.
  • A pixel-based approach for text separation based on [1] will be applied.
  • The results will be evaluated based on the ground truth data.


[1] Dutly, N., Slimane, F., & Ingold, R. (2019, September). Phti-ws: A printed and handwritten text identification web service based on fcn and crf post-processing. In 2019 International Conference on Document Analysis and Recognition Workshops (ICDARW) (Vol. 2, pp. 20-25). IEEE.


The project work will be supervised by Prof. Dr. Harald Sack, Mahsa Vafaie and Oleksandra Bruns, Information Service Engineering at Institute AIFB, KIT, in collaboration with FIZ Karlsruhe.


Keywords:

Machine Learning, CNN, pattern recognition


Pre-requisites:

Knowledge of Programming with Python.


Ausschreibung: Download (pdf)