Vietnamese Scenes Text Recognition Based on Deep Neural Networks

In recent years, there has been a proliferation of novel approaches pertaining to scene text recognition (STR) models. Although each model asserts to have expanded the technological frontier, their training and evaluation predominantly rely on English datasets, which are comparatively less challenging than datasets in other languages. An illustration of this can be seen in the Vietnamese language, which has a greater variety in its word structure. This intricacy can present a difficulty for STR models in effectively acquiring and making accurate predictions. The objective of our study is to ascertain the performance of models that have demonstrated proficiency on English scene text datasets when applied to Vietnamese scene text datasets. In order to arrive at a conclusion, we will assess various speech-to-text recognition (STR) models spanning from historical to contemporary iterations. These models will be evaluated using a dataset specific to the Vietnamese language. The models under evaluation encompass PARSeq, ABINet, ViTSTR, TRBA, and CRNN. While the aforementioned models have demonstrated strong performance on English scene text datasets, conducting tests on the VinText and BKAI datasets, which are Vietnamese Scene Text datasets, will enable us to ascertain the suitability of these models for training or fine-tuning purposes in the context of a challenging language.