KU Student’s AI Model Achieves 99.71% Accuracy In Devanagari Recognition

 

Kathmandu: Sahajraj Malla, a third-year undergraduate student at Kathmandu University, has developed an artificial intelligence (AI) model capable of accurately recognizing handwritten Devanagari script. His research paper on the model, named MallaNet, has been published in the prestigious international journal Nature.

According to the research paper, the model can recognize handwritten Devanagari characters with an accuracy of 99.71 percent. This development is considered significant for digitizing handwritten documents in Devanagari-based languages such as Nepali, Hindi, and Marathi, and for preserving cultural and linguistic heritage.

A key feature of MallaNet is its efficiency. While previous leading models use around 39 million parameters, MallaNet achieves comparable performance using 56.41 percent fewer parameters, with only 17 million. This allows the model to function effectively even on low-powered computing devices.

The model incorporates advanced techniques such as Homogeneous Filter Capsule (HFC) and Residual Blocks. These enable it to distinguish subtle differences between visually similar characters, such as ‘क’ and ‘ख’ or ‘ग’ and ‘ध’. The research used a dataset of 92,000 handwritten Devanagari character images across 46 categories.

This is the fifth research paper by Malla, who studies in the Department of Mathematics at Kathmandu University. His previous research has covered areas such as quantum machine learning and stock market forecasting. He has made the model’s code and results publicly available on Github to support further research.

The technology has potential applications beyond document digitization, including traffic number plate recognition, scanning old government records, and recognition of other regional languages

पछिल्लो अध्यावधिक: पुस १३, २०८२ १३:४४

टिप्पणीहरू