Contour-KNN Brahmi Segmentation (CKBS) and Two-Phase Enhanced Brahmi Recognition (TREBR) Methods for Automatic Brahmi Texts Labelling

Neha, Gautam (2024) Contour-KNN Brahmi Segmentation (CKBS) and Two-Phase Enhanced Brahmi Recognition (TREBR) Methods for Automatic Brahmi Texts Labelling. PhD thesis, Universiti Malaysia Sarawak.

	PDF Thesis PhD_NehaGautam.open -24 pages.pdf Download (877kB)
	PDF (Please get the password by email to repository@unimas.my , or call ext: 3942 / 3973 / 3933) Thesis PhD_NehaGautam.ftext.pdf Restricted to Registered users only Download (2MB) | Request a copy
	PDF Thesis PhD_NehaGautam.dsva.pdf Restricted to Repository staff only Download (229kB) | Request a copy

Abstract

Automatic word recognition problem can be solved using an optical character recognition (OCR) system. Few studies have been seen in the field of Brahmi word recognition especially identifying compound characters and words with good accuracy. However, existing Brahmi text recognition studies have primarily relied on local datasets, hampering the standardization of datasets. To address this, the study proposes a systematic dataset creation process that encompasses data collection, pre-processing, segmentation, data augmentation, recognition, storage, and labelling. The process is initiated with data collection from diverse online sources, yielding 217 text and word samples and 801 isolated characters and compound characters. However, these samples lack uniformity in text and word sizes. The subsequent phase focuses on character isolation from words and text, utilizing a novel segmentation approach as a crucial precursor to system training. A ContourKNN Brahmi Segmentation (CKBS) for character and compound character segmentation is introduced. Object detection identifies characters, including dots (.), and links them to their nearest left character using KNN. This approach greatly enhances segmentation, achieving an impressive 98.19% average accuracy. The segmentation approach generates 40 samples per class across 170 classes, with a 75:25 training-testing split (30 and 10 samples for training and testing, respectively). Furthermore, data augmentation techniques, including adjustments, deformations, blurring, translations, and noise introduction, are applied to enhance dataset quality and quantity. Data augmentation results in 180 training and 60 testing samples per class, improving both size and quality. Subsequently, a Two-Phase Enhanced Brahmi Recognition (TPEBR) is employed, distinguishing between global and local feature recognition. Various deep learning architectures are evaluated for classification, with resizing to meet specific input size requirements. SqueezeNet emerges as the most effective, achieving a minimal 0.237 loss and an exceptional 97.58% accuracy. It excels in precision, recall, and F1-score. In contrast, ResNeXt Small underperforms with higher loss and lower accuracy. Comparing the Two-Phase Enhanced Brahmi Recognition (TPEBR) to the existing approaches, the Two-Phase Enhanced Brahmi Recognition (TPEBR) achieves 97.58% accuracy, while the existing approaches records 80.20% and 90.24%. Recognized characters are then organized into folders according to their recognized class, and done labelling by using Brahmi Unicode, although this step does not impact performance of the system.

Actions (For repository members only: login required)

View Item