GARDEN: Generative Prior Guided Network for Scene Text Image Super-Resolution

Scene text image super-resolution (STISR) is a popular research topic due to its great potential for improving downstream recognition performance. Many recent STISR approaches have utilized recognition feedback to guide the reconstruction process. However, their effectiveness is often limited by inaccurate recognition feedback and insufficient use of visual priors. To address these challenges, we propose a novel GenerAtive pRior guiDEd Network, namely GARDEN, which surpasses existing practices by exploiting enriched generative priors for precise and reliable guidance towards STISR. Innovatively, GARDEN leverages a pre-trained Vision Transformer (ViT) as the generative style bank, which provides diverse image priors and further assists in generating reliable text priors. This allows the network to leverage prior information from both visual and semantic domains for the final reconstruction, leading to more efficient learning of both texture generation and text recovery. In addition, GARDEN introduces multi-scale sequential residual block (MS-SRB), a simple, efficient, and flexible structure for achieving the maximal utilization of generative priors. By leveraging enriched generative priors within a novel architecture design, GARDEN is better suited to encode, transfer, and reconstruct super-resolution text images than the best previous methods in terms of both fidelity and recognition accuracy, as shown in Fig. 1. Code will be publicly available.