Unveiling Cross-Linguistic Mastery: Advancing Multilingual Handwritten Numeral Recognition with Attention-driven Transfer Learning

Abstract

Abstract In the realm of data analysis and document processing, the recognition of handwritten numerals stands as a pivotal advancement. This contribution has steered transformative shifts in optical character recognition, historical handwritten document analysis, and postal automation. A persistent challenge in this arena is the recognition of handwritten digits across a spectrum of languages, each with its idiosyncrasies. We present an innovative paradigm to surmount this hurdle, transcending the confines of monolingual recognition. Unlike the status quo, which gravitates toward a narrow subset of languages, our method orchestrates a comprehensive solution spanning 12 distinct languages, deftly navigating linguistic intricacies. The catalyst for this efficacy is transfer learning, amplifying image quality and recognition acumen. Emboldening this framework is an ingenuity-charged attention-based module that refines precision. Our rigorous experimentations substantiate quantum leaps in image quality and the prowess of linguistic and numeral recognition. Notably, we unearth significant accuracy strides, eclipsing 2% enhancements in specific languages vis-à-vis antecedent methodologies. This endeavor epitomizes a sturdy, economically sound avenue, unshackling multilingual handwritten numeral recognition to an expansive spectrum of languages.