Jawi sub-word recognition system using window-based segmentation-free approach / Roslim Mohamad
The presence of features such as cursive, diversity of writing styles and sizes of characters in a Jawi text, ligature, and vertical overlapping make the recognition of Jawi handwritten text to be difficult. For the recognition system based on model development using windows, the presence of such fe...
Saved in:
| Main Author: | |
|---|---|
| Format: | Book Section |
| Language: | English |
| Published: |
Institute of Graduate Studies, UiTM
2018
|
| Subjects: | |
| Online Access: | http://ir.uitm.edu.my/id/eprint/20699/1/ABS_ROSLIM%20MOHAMAD%20TDRA%20VOL%2013%20IGS_18.pdf http://ir.uitm.edu.my/id/eprint/20699/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The presence of features such as cursive, diversity of writing styles and sizes of characters in a Jawi text, ligature, and vertical overlapping make the recognition of Jawi handwritten text to be difficult. For the recognition system based on model development using windows, the presence of such features cause the resulting model to be less consistent, where it produces a different sequence of primitive structures of words/ sub-words from the same lexicon. To overcome the inconsistency model problem, a handwritten Jawi text recognition system based on a subword model has been developed. The proposed modeling technique which is known as Selection Segmentation-Free (SSF) separates core and connection structure of a sub-word into a different window. The resulting window will go through a selection process to determine the windows that will be used to represent the sub-word model. In order to increase accuracy and efficiency of the representation feature, two categories of features which are known as primary and secondary features were extracted from each of the selected windows. Primary feature were extracted using Window Code Representation (WCR) technique from main structure. Secondary feature for supporting the primary feature were extracted from dot and main structure. For the experiment purposes, a total of 1200 sub-words of 80 lexicons were used. Each lexicon is randomly selected and divided into three sets. Three experiments to evaluate the performance of SSF, WCR and combination of primary and secondary feature techniques were conducted. The three techniques are combined to represent the proposed system and compared with the comparison system introduced by Remon (2009). Comparison result shows that the recognition rate of proposed system (84.8%) is better than comparison system (79.1%). |
|---|