Parametric loss-based super-resolution for scene text recognition

Scene text image super-resolution (STISR) is regarded as the process of improving the image quality of low-resolution scene text images to improve text recognition accuracy. Recently, a text attention network was introduced to reconstruct high-resolution scene text images; the backbone method involved the convolutional neural network-based and transformer-based architecture. Although it can deal with rotated and curved-shaped texts, it still cannot properly handle images containing improper-shaped texts and blurred text regions. This can lead to incorrect text predictions during the text recognition step. In this study, we propose the application of multiple parametric regularizations and parametric weight parameters to the loss function of the STISR method to improve scene text image quality and text recognition accuracy. We design and extend it into three types of methods: adding multiple parametric regularizations, modifying parametric weight parameters, and combining parametric weights and multiple parametric regularizations. Experiments were conducted and compared with state-of-the-art models. The results showed a significant improvement for every proposed method. Moreover, our methods generated clearer and sharper edges than the baseline with a better-quality image score.