Show-through phenomena have always been a challenging issue in color-document image processing, which is widely used in various fields such as finance, education, and administration. Existing methods for processing color-document images face challenges, including dealing with double-sided documents with show-through effects, accurately distinguishing between foreground and show-through parts, and addressing the issue of insufficient real image data for supervised training. To overcome these challenges, this paper proposes a self-supervised-learning-based method for removing show-through effects in color-document images. The proposed method utilizes a two-stage-structured show-through-removal network that incorporates a double-cycle consistency loss and a pseudo-similarity loss to effectively constrain the process of show-through removal. Moreover, we constructed two datasets consisting of different show-through mixing ratios and conducted extensive experiments to verify the effectiveness of the proposed method. Experimental results demonstrate that the proposed method achieves competitive performance compared to state-of-the-art methods and can effectively perform show-through removal without the need for paired datasets. Specifically, the proposed method achieves an average PSNR of 33.85 dB on our datasets, outperforming comparable methods by a margin of 0.89 dB.

When documents are scanned and converted into digital images, they often contain unwanted and unevenly superimposed show-through layers, which directly impact the performance of other vision-related tasks. The show-through removal is significant in image processing applications. Show-through refers to the presence of transparency or uneven materials in document images during the scanning or capturing process, resulting in a visible transparency effect. In this study, we propose a two-stage framework called document show-through removal generative adversarial networks that utilizes conditional generative adversarial networks to restore document images affected by show-through phenomena. The coarse stage consists of two parallel generative architectures (Gremove and Gpredict). We employ a residual block structure between the encoder and decoder of Gremove to extract deeper features. In Gpredict, we utilize a deep residual attention block based on an attention mechanism, substituting the traditional convolutional layer. Finally, the results are fused in the refinement stage (Grefine) to generate higher-quality document images. Experimental results demonstrate that our method not only effectively removes show-through information but also significantly enhances the clarity of foreground content. Compared to state-of-the-art methods, our approach achieves a better performance, such as 0.47 dB in peak signal-to-noise ratio. The source code is available at https://github.com/GongJiaXin694/DSR-GAN.