An algorithm for constructing character data of the Uyghur-style Mongolian script for machine vision recognition

Abstract

This paper proposes a visual data construction algorithm for Uighur-style Mongolian characters aimed at machine vision recognition. Building upon traditional character image preprocessing and CNN recognition frameworks, the algorithm incorporates structure-aware segmentation, skeleton topological encoding, annotation correction, and structure-preserving expansion techniques to construct training samples that include images, labels, topological vectors, and quality scores. Experimental results show that the model achieves an accuracy of 91.8%, representing a 7.2 percentage point improvement over the original manually cropped data. The Macro-F1 score increased from 0.831 to 0.912, and the recall for low-frequency categories rose from 0.752 to 0.876.