Research map / Image generation / Image restoration
End-to-End Unsupervised Document Image Blind Denoising
Mehrdad J. Gangeh, Marcin Plata, Hamid R. Motahari Nezhad, Nigel P Duffy
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add classifier C to the embedder (ablation) | Proposed MoE cycle-GAN · Dataset I (watermarked scientific pages) | PSNR: 37.99 -> 38.33 | Improved |
| add classifier C to the embedder (ablation) | Proposed MoE cycle-GAN · Dataset I (watermarked scientific pages) | Max deterioration (%): 14.73 -> 9.93 | Improved |
| add classifier C to the embedder (ablation) | Proposed MoE cycle-GAN · Dataset I (watermarked scientific pages) | Averaged deterioration (%): 2.54 -> 2.55 | No clear effect |
| add deep mixture of experts with embedder to cycle-GAN | cycle-GAN · Dataset I (watermarked scientific pages) | Averaged deterioration (%): 7.85 -> 2.55 | Improved |
| add deep mixture of experts with embedder to cycle-GAN | cycle-GAN · Dataset I (watermarked scientific pages) | PSNR: 36.52 -> 38.33 | Improved |
| add deep MoE to cycle-GAN | cycle-GAN · Dataset II (Tobacco800 noisy pages) | Averaged OCR improvement (%): 5.94 -> 7.2 | Improved |
| add deep MoE to cycle-GAN | cycle-GAN · Dataset III (in-house noisy pages) | Averaged OCR improvement (%): 6.30 -> 9.52 | Improved |
| replace sequential cycle-GAN training with unified MoE model | cycle-GAN (Sequential) · Dataset II (Tobacco800 noisy pages) | Averaged OCR improvement (%): 6.03 -> 7.2 | Improved |
| replace sequential cycle-GAN training with unified MoE model | cycle-GAN (Sequential) · Dataset III (in-house noisy pages) | Averaged OCR improvement (%): 7.10 -> 9.52 | Improved |
| use unsupervised unified model instead of supervised REDNet | REDNet · Dataset I (watermarked scientific pages) | Averaged deterioration (%): 1.84 -> 2.55 | Hurt |
| use unsupervised unified model instead of supervised DE-GAN | DE-GAN · Dataset I (watermarked scientific pages) | Averaged deterioration (%): 3.74 -> 2.55 | Improved |
| clean pages with proposed model before OCR | ABBYY OCR · Dataset I (watermarked scientific pages) | Averaged deterioration (%): 7.71 -> 2.55 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DE-GAN — DeGAN: Data-Enriching GAN for Retrieving Representative Samples from a Trained Classifier (AAAI 2020) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Unsupervised Image Restoration Using Partially Linear Denoisers (TPAMI 2021 · similarity 0.85)
- Unpaired Learning of Deep Image Denoising (ECCV 2020 · similarity 0.85)
- End-to-End Unpaired Image Denoising with Conditional Adversarial Networks (AAAI 2020 · similarity 0.85)
- Fully Unsupervised Diversity Denoising with Convolutional Variational Autoencoders (ICLR 2021 · similarity 0.85)
- IDR: Self-Supervised Image Denoising via Iterative Data Refinement (CVPR 2022 · similarity 0.84)
- Efficient Deep Image Denoising via Class Specific Convolution (AAAI 2021 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
