Handwritten text recognition has been an active research topic within computer vision division. Existing deep-learning solutions are practical; however, recognizing Vietnamese handwriting has shown to be a challenge with the presence of extra six distinctive tonal symbols and extra vowels. Vietnam is a developing country with a population of approximately 100 million, but has only focused on digitalization transforms in recent years, and so Vietnam has a significant number of physical documents, that need to be digitized. This digitalization transform is urgent when considering the public health sector, in which medical records are mostly still in hand-written form and still are growing rapidly in number. Digitization would not only help current public health management but also allow preparation and management in future public health emergencies. Enabling the digitalization of old physical records will allow efficient and precise care, especially in emergency units. We proposed a solution to Vietnamese text recognition that is combined into an end-to-end document-digitalization system. We do so by performing segmentation to word-level and then leveraging an artificial neural network consisting of both convolutional neural network (CNN) and a long short-term memory recurrent neural network (LSTM) to propagate the sequence information. From the experiment with the records written by 12 doctors, we have obtained encouraging results of 6.47% and 19.14% of CER and WER respectively.
Phung, Tuan Minh; Dinh, Minh Ngoc; Dang, Duy Pham Thien; Van, Hoang Minh Tu; and Thwaites, C. Louise, "A Machine Learning-based Approach to Vietnamese Handwritten Medical Record Recognition" (2020). ACIS 2020 Proceedings. 22.