ZZZHANG-jx/Recommendations-Document-Image-Processing

What it solves

Photographed documents often suffer from shadows, curling, blur, uneven lighting, or poor contrast — problems that make them hard to read or process automatically. This project gathers the latest research on fixing those issues, covering tasks like removing shadows, flattening curled pages, sharpening blur, and improving overall appearance so a photo of a document looks as clean as a scanned or digital PDF.

How it works

This is not a software tool but a well-organised literature collection. It groups papers, datasets, and state-of-the-art (SOTA) results into task-specific sections. Each section lists relevant papers with venues and links to code where available, corresponding datasets with sizes and download links, and a SOTA table comparing recent methods on benchmark metrics like SSIM, PSNR, RMSE, and more. The sections cover registration (aligning two images of the same document), all-in-one restoration (a single model handling multiple degradations), appearance enhancement, deshadowing, dewarping, deblurring, binarization, and handwritten text erasure. The repo also includes example images for several datasets.

Who it's for

Researchers, students, or engineers working on document image processing or OCR pipelines. If you want to know which methods are state-of-the-art, which datasets are commonly used, or where to get code for a specific technique, this collection saves hours of literature searching.

Highlights

• Covers eight major document image processing tasks, plus a section on related apps and tools.
• Each task section lists recent papers (2018–2026) from top venues like CVPR, ICCV, and ACM MM, with links to code repositories where available.
• Included datasets vary from synthetic to real-world, with download links and example images.
• SOTA tables give quick comparisons on public benchmarks, e.g., the DocUNet and RealDAE datasets for appearance enhancement, or Kligler, Jung, OSR, RDD, and SD7K for shadow removal.
• A unified restoration model section (All in One) highlights recent efforts like DocRes and diffusion-based approaches that handle multiple degradations in a single model.