Scholarlink

<p>Document images often suffer from blurring, noise, and shadows, which can seriously degrade their visual quality and hinder the effectiveness of enhancement models. However, most existing methods struggle with performance degradation in rotated watermarks and slow processing speed. To address these issues, we present DeepDiff, a novel document enhancement model. DeepDiff features a Quaternion Attention Mechanism (QAM) that uses quaternion multiplication to effectively capture rotational features, enabling the model to better handle rotated watermarks. In addition, DeepDiff employs a divide-and-conquer parallel design, allowing different parts of an image to be processed simultaneously on multiple GPUs, which accelerates generation speed. Experiments on the Denoising Dirty Documents dataset show that DeepDiff achieves a PSNR of 26.40 and an SSIM of 0.9688. The code is available at <a href="https://github.com/Richard-153/code.">https://github.com/Richard-153/code.</a></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DeepDiff: Document enhancement via enhanced parallel diffusion model

  • 期刊论文

  • 发表日期：

    2026年2月26日

查看全文
  • Guohao Cui,
  • Cihui Yang

摘要

Document images often suffer from blurring, noise, and shadows, which can seriously degrade their visual quality and hinder the effectiveness of enhancement models. However, most existing methods struggle with performance degradation in rotated watermarks and slow processing speed. To address these issues, we present DeepDiff, a novel document enhancement model. DeepDiff features a Quaternion Attention Mechanism (QAM) that uses quaternion multiplication to effectively capture rotational features, enabling the model to better handle rotated watermarks. In addition, DeepDiff employs a divide-and-conquer parallel design, allowing different parts of an image to be processed simultaneously on multiple GPUs, which accelerates generation speed. Experiments on the Denoising Dirty Documents dataset show that DeepDiff achieves a PSNR of 26.40 and an SSIM of 0.9688. The code is available at https://github.com/Richard-153/code.