Learning-based image mapping for degraded documents on E-paper display

Abstract

With the widespread use of E-paper technology, numerous documents are being digitized and displayed on E-paper screens. However, the display quality of degraded document images on E-paper often suffers from a lack of detail. To address this challenge, we introduce a mapping model that converts color images into E-paper display images. This model leverages U-Net++ as its backbone, integrating residual connectivity and dual attention modules. Given the presence of varying stroke thicknesses in document images, a fixed-size convolutional kernel is insufficient. Therefore, we propose multi-branch channels and spatial attention modules (MCSAM), which combines the selective kernel network (SKNet) with a spatial attention mechanism to adaptively select the appropriate convolutional kernel size based on font size. To demonstrate its effectiveness, we tested the mapped images on a custom E-paper display platform. Experimental results highlight the superior performance of our proposed method.