Please use this identifier to cite or link to this item:
https://ptsldigital.ukm.my/jspui/handle/123456789/513445Full metadata record
| DC Field | Value | Language |
|---|---|---|
| dc.contributor.advisor | Khairuddin Omar, Prof. Dr. | |
| dc.contributor.author | Anton Heryanto Hasan (P52862) | |
| dc.date.accessioned | 2023-10-16T04:36:45Z | - |
| dc.date.available | 2023-10-16T04:36:45Z | - |
| dc.date.issued | 2019-10-09 | |
| dc.identifier.other | ukmvital:130949 | |
| dc.identifier.uri | https://ptsldigital.ukm.my/jspui/handle/123456789/513445 | - |
| dc.description | The digitisation of Jawi handwritten manuscript is very important to allow efficient archiving and retrieving of the original documents and increasing the availability of the content. However, Jawi handwriting recognition is a challenging task. The problems and challenges in Jawi handwriting recognition are inherited from Arabic script which includes the use of cursive, a large variety of writing styles, ligature, overlapping characters and large lexicon size due to varieties of rules and dialects. This is further compounded by the often low quality of the manuscript images. The existence of disconnect characters introduces the sub word problem which is the inter word space that is sometimes bigger than the space between the words. The performance of previous Jawi handwriting subword recogniser is still considered subpar. The multiple independent components used are hard to optimize and the improvement of one component does not necessarily translate into better overall performance. The segmentationbased recognition approach tends to cause the loss of information in character segmentation and result in missclassification. Segmentationfree approach is only usable for a limited lexicon and it is also unable to handle the large varieties of sub word class. The state of the art Jawi handwriting subword recognition uses trace transform object signature features which are invariance regarding size or rotation. Despite its potential, the circular natures of object signature features produce subpar performance when combined with machine learning classifier. The features are handcrafted using feature engineering approach which is quite tedious and sub optimum to find the best features. This research proposed deep learning based Jawi handwriting subword recogniser were whole component integrated in a big network. The parameters of each component are adjusted in endtoend in training, from raw input to the last output to improve the overall system performance. Using the the high representational capacity of deep learning, raw images of subword is implicitly segmented into sequence of characters and each character is recognized by position dependent multiple character classifier into Unicode. It consider lexiconfree approach as lexiconguide requirement for subword recognition became optional. The trace transform feature learning improves the robustness of the trace transform feature by automatic adjust parameters and select the best feature which is integrated with classifiers to solve certain task. Its singlelayer performance is better compared to singlelayer and three layers of convolutional neural network which is the state of the art in feature learning. The combination of global feature of trace transform with local features convolutional neural network produce more robust feature which further improves the Jawi handwriting recognition performance. The proposed Jawi handwriting recogniser is significantly outperformed the state of the art Jawi handwriting recogniser recognition performance with 52.17% percent improvement.,Ph.D | |
| dc.language.iso | eng | |
| dc.publisher | UKM, Bangi | |
| dc.relation | Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat | |
| dc.rights | UKM | |
| dc.subject | Universiti Kebangsaan Malaysia -- Dissertations | |
| dc.subject | Dissertations, Academic -- Malaysia | |
| dc.subject | Jawi handwriting | |
| dc.subject | Neural network | |
| dc.subject | Trace transform network | |
| dc.subject | Neural networks (Computer science) | |
| dc.title | Jawi handwriting recognition using trace transform network and convolutional neural network with multiple character classifier | |
| dc.type | Theses | |
| dc.format.pages | 207 | |
| Appears in Collections: | Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat | |
Files in This Item:
| File | Description | Size | Format | |
|---|---|---|---|---|
| ukmvital_130949+Source01+Source010.PDF Restricted Access | 3.07 MB | Adobe PDF |  View/Open |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.