FAUNet: A Fuzzy-Attention U-Net for Diffusion-Based Persian Text Image Super-Resolution

The accurate enhancement of text images is a critical challenge in computer vision, particularly for languages such as Persian that exhibit complex writing structures, cursive connections, and fine-grained diacritical marks. Traditional super-resolution approaches often fail to preserve these delicate textual details. Here, diffusion model is adopted for text image super-resolution. The U-Net framework of this method is enhanced by incorporating fuzzy logic and attention mechanism (named FAUNet) to address mentioned problems. At the bottleneck of the network, a fuzzy layer is employed to softly model uncertainties and boundary variations, while a spatial channel attention block adaptively emphasizes crucial regions of the image. Together, these components strengthen the network’s capacity to capture structural dependencies and semantic details essential for text clarity. The proposed model is rigorously evaluated on two large-scale Persian text datasets: IR-LPR that comprising vehicle license plate images, and IDPL-PFOD۲ that is a dataset of printed Persian text. Experimental results show that FAUNet outperforms state-of-the-art methods achieving improvements in PSNR, SSIM, and MS-SSIM metrics. These improvements not only contribute to higher visual quality but also hold strong potential for downstream applications such as optical character recognition (OCR), license plate recognition, and digital document restoration in low-quality imaging conditions.

کلیدواژه ها:

Super-Resolution ، Deep learning ، Diffusion model ، Fuzzy layer ، Attention mechanism