Scene Text Image Super-Resolution Guided by Frequency Domain Enhancement and Feature Refinement

Scene text image super-resolution (STISR) is a method to improve the quality of low-resolution text images, aiming to improve the readability and recognition accuracy of text by increasing the resolution of images. The current STISR method ignores the problem that high-frequency features such as edges and textures are difficult to extract when extracting low-resolution image features, as well as the noise generated during feature processing. However, these high-frequency features are crucial in the scene text image super-resolution task. To address this problem, we propose a frequency enhancement and feature refinement guided scene text image super-resolution (FEFR-STISR) model. Based on the traditional STISR method, this model uses high-pass filters and feature refinement modules to perform frequency domain enhancement and feature refinement, highlighting high-frequency features such as edges and textures while reducing the impact of noise. Among them, the function of the image feature branch is to extract the image features of the low-resolution image, and the function of the Text Probability (TP) generation branch is to extract the TP features in the image. Finally, the extracted image features are enhanced in the frequency domain and then fused and refined with the TP features in the cyclic residual module and finally upsampled into a super-resolution image. Experiments on the TextZoom dataset show that our FEFR-STISR model exhibits superior performance.