Enhanced edge detection and restoration methods for stroke width transform algorithm
No Thumbnail Available
Date
2018
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Teknologi Malaysia
Abstract
Text localisation in images has gained widespread interests. A notable work, which is the Stroke Width Transform (SWT) has been attracting much interests due to its simplicity and efficiency. However, the problems of imperfect edge map and stroke-liked objects have limited the SWT performance. Imperfect edge map can affect text localisation ability while stroke-liked objects will cause faulty judgement to real text. This research attempted to solve the problems by enhancing the localisation capability in three different phases in the SWT. First, an Edge Restoration (ER) algorithm for restoring broken edges based on the characteristic of the SWT for better edge map production was introduced. Next, to reduce SWT dependency on image colour based on the edge map inherited from the previous stage, Stroke Width Map (SWM) generation algorithm was developed. Finally, a Text Filtering (TF) algorithm distinguishing stroke-liked objects in the image based on the features obtained from the previous stages was developed. Two experiments were conducted to evaluate the performance of the proposed algorithms. The first experiment evaluated the ER algorithm on the completeness of the generated edge map whereas the second experiment evaluated the text localisation performance of SWT after implementing SWM generation and TF algorithms. Experiment results showed that all algorithms have successfully improve the localisation performance of SWT. Firstly, the ER algorithm has the ability to recover broken edge structure and identify noise in the edge map. Secondly, the SWM generation algorithm generated a more accurate SWM with less interference from text colour and excess edges. Lastly, the TF algorithm distinguished between real text and the stroke-liked non-text objects. These results indicate that the ER, SWM generation and TF algorithms have overcome the imperfect edge map and stroke-liked objects problems, and improve the localisation capability in SWT.
Description
Thesis (PhD. (Computer Science))
Keywords
Optical character recognition, Optical pattern recognition—Data processing, Image processing—Computer programs