Semantic prior guided text image super-resolution with dynamic feature fusion

Text image super-resolution has received widespread attention for its ability to improve text recognition accuracy in text images. The key aspects of text image super-resolution are extracting semantic priors for guidance and fusing semantic priors with visual features, aiming to restore stroke details in low-resolution text images. Therefore, high-quality semantic priors and effective feature fusion are particularly important. However, most recent research has focused on complex super-resolution model architectures, often neglecting the effectiveness of semantic priors and feature fusion methods. Within this study, we introduce a novel approach for enhancing the resolution of text-based images with semantic prior correction for low-resolution images. Specifically, to improve the effectiveness of semantic priors, we introduce a semantic prior correction module that learns the data distribution of high and low-quality semantic priors, optimizing the extracted semantic priors to achieve more efficient prior guidance. Furthermore, to enhance the deep fusion interaction process between semantic priors and visual features, we design a dynamic feature fusion module that employs attention mechanisms to achieve deep dynamic interaction between the two types of features. Our experiments demonstrate that our approach surpasses current leading techniques in terms of both numerical metrics and visual assessment.