An enhanced diffusion-based network for efficient stamp removal

Abstract

Current diffusion models excel in computer vision tasks, but stamp removal from documents remains challenging, especially when stamps are light-colored and blend with text. Existing methods struggle to preserve background text and rely heavily on the training set, excelling in either text or table stamp removal, but not both. To address these problems, we propose an enhanced diffusion-based stamp removal model using a Spatial Attention Mechanism and a Simulate Rectified Linear Unit. Spatial Attention Mechanism combines the spatial transformation capabilities of the Spatial Transformer Network with the feature extraction of the Convolutional Block Attention Module for higher-quality images. Simulate Rectified Linear Unit mimics neuronal signal transmission in the human brain, enhancing feature extraction. Our diffusion model achieved a PSNR of 44.7, SSIM of 0.99, and RMSE of 3.47 on the stamp dataset, and performed optimally on the denoising-dirty-documents, CLWD, and DIBCO 2017 datasets. It also attained the highest PSNR of 26.8 on the DIBCO 2013 dataset, with other metrics close to the best. Code is available at https://github.com/GuohaoCui/DiffusionModel.