Scene Text Recognition Based on Improved CRNN

Abstract

Text recognition is an important research topic in computer vision. Scene text, which refers to the text in real scenes, sometimes needs to meet the requirement of attracting attention, and there is the situation such as deformation. At the same time, the image acquisition process is affected by factors such as occlusion, noise, and obstruction, making scene text recognition tasks more challenging. In this paper, we improve the CRNN model for text recognition, which has relatively low accuracy, poor performance in recognizing irregular text, and only considers obtaining text sequence information from a single aspect, resulting in incomplete information acquisition. Firstly, to address the problems of low text recognition accuracy and poor recognition of irregular text, we add label smoothing to ensure the model’s generalization ability. Then, we introduce the smoothing loss function from speech recognition into the field of text recognition, and add a language model to increase information acquisition channels, ultimately achieving the goal of improving text recognition accuracy. This method was experimentally verified on six public datasets and compared with other advanced methods. The experimental results show that this method performs well in most benchmark tests, and the improved model outperforms the original model in recognition performance.

Keywords

• CRNN
• text recognition
• label smoothing
• language model
• deep learning