DocTrKan: Document Image Enhancement Based on KAN-Enhanced Transformer Networks

Document images serve as a medium for transmitting information; however, they are influenced by numerous environmental factors during capture, complicating the identification and processing of their content. In the digital era, it is crucial to eliminate noise from document images to enhance readability. To address this challenge, we propose a novel transformer-based document image enhancement algorithm utilizing KAN enhancement, designed to improve both machine-printed and handwritten document images in an end-to-end manner. The encoder operates directly on pixel blocks with positional information, bypassing traditional convolutional layers for feature extraction. A KAN network is incorporated into the encoder to enhance the model's expressive power, reduce parameter count, and improve optimization efficiency. A decoder is employed to reconstruct the clean image. In testing on the DIBCO dataset, DocTrKan surpasses traditional transformer and threshold-based methods in enhancement performance.