Improving Classification of Scanned Document Images using a Novel Combination of Pre-Processing Techniques

Mahajan, Arpana Dipak; Karuppasamy, Selvakuberan; Lakshminarayanan, Subhashini

doi:10.1109/icaaic56838.2023.10140911

Cited by 2 publications

(1 citation statement)

References 17 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The study highlighted the importance of optimising system parameters and integrating semantic models to enhance word prediction across different languages. [5] Research underscoring the significance of feature extraction techniques in image processing was studied which emphasised the role of preprocessing methods such as binarization, thresholding, resizing, and normalisation. The study explored diverse feature extraction methods, including statistical features, global transformations, and local geometric features, providing insights into their applications in pattern recognition systems.…”

Section: Related Workmentioning

confidence: 99%

Unveiling the Past: A Holistic Approach to Rescuing Historical Texts through Advanced Image Analysis and AI

Kadam,

Jadhav,

Govilkar

et al. 2024

Preprint

View full text Add to dashboard Cite

In recent years, more and more administrative and legal paperwork has been done using computers because it's efficient. But in the past, when technology wasn't as good, there were problems with printing documents. This meant that text could be hard to read, the ink might not be consistent, and documents could get damaged over time. This caused a lot of important information, especially in ancient scripts, to be lost because it couldn't be digitised. Our research is all about solving this problem. We want to find a way to automatically get accurate information from pictures of these old documents. We employ a multi-step approach to enhance text recovery. Firstly, we utilise adaptive Gaussian thresholding to improve image clarity by removing excess ink or stains. Next, we apply optical character recognition (OCR) using advanced systems like Easy OCR and Tesseract, followed by thorough database validation for result accuracy. But sometimes, even with the special software, we can't see all the words because they're too faded or missing. To overcome this, we employ Natural Language Processing (NLP) techniques with Happy Transformer, a tool based on Hugging Face's Transformer Library. This enables us to predict and reconstruct missing letters or words. After prediction and cross-referencing with the database, the identified words are seamlessly integrated into the text, ensuring data retrieval. Our research is important for preserving historical and legal documents that might otherwise be lost forever.

show abstract

Section: Related Workmentioning

confidence: 99%