Presentation | 2023-11-17 Diffusion-based Geometric Unwarping and Illumination Correction for Document Images Sota Imahayashi, Guoqing Hao, Satoshi Iizuka, Kazuhiro Fukui, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | This study proposes a method to improve the visibility of document images by correcting distortions and re-illuminating them using a latent diffusion model. Document images often suffer from folds, tilt, and shadows. Such distortions and shadows significantly diminish image visibility, posing a challenge for optical character recognition (OCR) tasks. Recent methods using convolutional neural networks have attempted to correct distortions and re-illuminate, but image transformations involving geometric deformations are difficult, and it is still difficult to correct distortions and re-illuminate with high accuracy. In this study, we propose a method for correcting distortion and re-illumination of document images using a latent diffusion model to solve these problems. In the proposed method, a degraded image containing distortions and shadows is transformed from pixel space to latent space and concatenated with Gaussian noise. Then, the process of removing the noise is repeated by a denoising network to generate a latent representation with completely removed noise. Finally, the generated latent representation is converted to pixel space, and the output image is obtained after distortion correction and re-illumination. The diffusion model has high generative capacity and is trained by maximum likelihood estimation, allowing for the generation of a wide variety of data and stable training. It can also recover the remainder from partial information and conditionally control the generation, making it possible to generate images with distortion correction and re-illumination conditional on a degraded image. This method is expected to achieve high-quality results in a wide variety of real-world document images, and has a wide range of potential applications. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | Document images / Geometric unwarping / Illumination correction / Latent diffusion model |
Paper # | PRMU2023-36 |
Date of Issue | 2023-11-09 (PRMU) |
Conference Information | |
Committee | PRMU / IPSJ-CVIM / IPSJ-DCC / IPSJ-CGVI |
---|---|
Conference Date | 2023/11/16(2days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | |
Chair | Kunio Kashio(NTT) |
Vice Chair | Takuya Funatomi(NAIST) / Go Irie(Tokyo Univ. of Science) |
Secretary | Takuya Funatomi(Tokyo Inst. of Tech.) / Go Irie(Riken) |
Assistant | Kei Shimonishi(Kyoto Univ.) / Kensho Hara(AIST) |
Paper Information | |
Registration To | Technical Committee on Pattern Recognition and Media Understanding / Special Interest Group on Computer Vision and Image Media / Special Interest Group on Digital Contents Creation / Special Interest Group on Computer Graphics and Visual Informatics |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | Diffusion-based Geometric Unwarping and Illumination Correction for Document Images |
Sub Title (in English) | |
Keyword(1) | Document images |
Keyword(2) | Geometric unwarping |
Keyword(3) | Illumination correction |
Keyword(4) | Latent diffusion model |
1st Author's Name | Sota Imahayashi |
1st Author's Affiliation | University of Tsukuba(Univ. of Tsukuba) |
2nd Author's Name | Guoqing Hao |
2nd Author's Affiliation | University of Tsukuba(Univ. of Tsukuba) |
3rd Author's Name | Satoshi Iizuka |
3rd Author's Affiliation | University of Tsukuba(Univ. of Tsukuba) |
4th Author's Name | Kazuhiro Fukui |
4th Author's Affiliation | University of Tsukuba(Univ. of Tsukuba) |
Date | 2023-11-17 |
Paper # | PRMU2023-36 |
Volume (vol) | vol.123 |
Number (no) | PRMU-266 |
Page | pp.pp.113-118(PRMU), |
#Pages | 6 |
Date of Issue | 2023-11-09 (PRMU) |