DPDTRN: a dynamic pixel-level difficulty-aware texture reconstruction network for document super-resolution

Existing super-resolution (SR) reconstruction methods have achieved notable success in improving the texture details in low-resolution (LR) document images. However, most methods fail to consider variations in texture reconstruction difficulty across different regions, treating the difficulty of reconstructing textures for each object in a document image as equal and applying the same approach. Consequently, when faced with high-difficulty restoration textures (HDRT) such as character adhesion or background-text aliasing, they often exhibit semantic deviations and block artifacts. Some methods partition LR images into patches of equal size and discriminate based on the patch’s reconstruction difficulty, but fail to distinguish the varying reconstruction difficulties among different textures within each patch. Therefore, we propose a dynamic pixel-level difficulty-aware texture reconstruction network (DPDTRN). DPDTRN comprises a pixel-level difficulty-aware module that adaptively perceives HDRT across all objects within document image, a dedicated HDRT reconstruction module, and a loss function that facilitates differentiated reconstruction according to the difficulty level of HDRT. Compared to state-of-the-art (SOTA) approaches, DPDTRN achieves superior performance with 2.05 dB gain in peak signal-to-noise ratio (PSNR) and 0.014 gain in structure similarity index measure (SSIM) on the complex document image dataset Text330. Moreover, it produces more accurate reconstructions for small characters prone to semantic transfer such as ’l’, ’i’, ’O’, and characters with aliased textures overlapping the background, additionally exhibiting reduced block artifacts on dark backgrounds. Meanwhile, we demonstrate the robustness of DPDTRN on natural image dataset DIV2K, where it showcases strong performance in HDRT. The code is available at https://github.com/gpeng5021/DPDTRN.