Uyghur Text Recognition Based on the SVTR Network

To address the lack of image datasets for Uyghur text recognition, text image synthesis techniques were used to construct the Uyghur text image datasets; data enhancement techniques such as adding noise, blurring, and brightness changes were used to enhance some of the image in the dataset to varying degrees to further enhance the natural scene features in the constructed dataset; finally, 99.9% text recognition accuracy was obtained on the constructed dataset using the SVTR network with the addition of the Uyghur text dictionary, with a 21.9% increase in accuracy compared to the transformed Latin CRNN network for the same dataset. Compared with other Uyghur text recognition methods under previous machine learning frameworks, the method in this paper can batch process data, automatically extract features to improve efficiency, and have different degrees of improvement in recognition results from a few points to twenty points. The experimental results show that the recognition method using synthetic images and data enhancement techniques under this paper's deep learning network framework can better achieve the recognition task of Uyghur text and achieve better experimental results.