Research map / Adversarial robustness and security / Certified and empirical robustness
Total Deep Variation: A Stable Regularization Method for Inverse Problems
Erich Kobler, Alexander Effland, Karl K. Kunisch, Thomas Pock
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 |
|---|---|---|---|
| use identity potential psi(x)=x instead of ln cosh(x) (ablation) | TDV_3^3 · BSDS68 gray denoising sigma=25 | PSNR: 29.3596 -> 29.3722 | Improved |
| use identity potential psi(x)=x instead of 1/2 log(1+x^2) (ablation) | TDV_3^3 · BSDS68 gray denoising sigma=25 | PSNR: 29.3662 -> 29.3722 | Improved |
| train TDV individually per noise level instead of one model trained for sigma=25 | TDV_3^3 · Urban100 gray denoising sigma=50 | PSNR: 26.94 -> 27.04 | Improved |
| train TDV individually per noise level instead of one model trained for sigma=25 | TDV_3^3 · Urban100 gray denoising sigma=15 | PSNR: 32.66 -> 32.91 | Improved |
| train TDV individually per noise level instead of one model trained for sigma=25 | TDV_3^3 · BSDS68 gray denoising sigma=25 | PSNR: 29.37 -> 29.37 | No clear effect |
| increase number of blocks and scales of the TDV regularizer (ablation) | TDV · BSDS68 gray denoising sigma=25 | expected PSNR | Improved |
| increase depth S beyond 10 (ablation) | TDV · BSDS68 gray denoising sigma=25 | expected PSNR | No clear effect |
| stop gradient flow at optimal stopping time T*=0.172 rather than running longer (ablation) | TDV_3^3 · BSDS68 gray denoising sigma=25 | PSNR: T*=0.172 | Improved |
| use TDV_3^3 instead of FOCNet | TDV · gray denoising parameter count | trainable parameters: 53,513,120 -> 387,394 | Improved |
| use TDV_{3,25}^3 instead of FFDNet | TDV · Kodak24 color denoising sigma=50 | PSNR: 28.98 -> 29.44 | Improved |
| use TDV_{3,25}^3 instead of FFDNet | TDV · McMaster color denoising sigma=15 | PSNR: 34.66 -> 34.55 | Hurt |
| use TDV_3^3 instead of OISR-LF-s | TDV · Set14 super-resolution x2 | PSNR: 33.62 -> 33.35 | Hurt |
Closest papers by meaning
- Total Deep Variation for Linear Inverse Problems (CVPR 2020 · similarity 0.94)
- Adversarial Regularizers in Inverse Problems (NeurIPS 2018 · similarity 0.93)
- Weakly Convex Regularisers for Inverse Problems: Convergence of Critical Points and Primal-Dual Optimisation (ICML 2024 · similarity 0.89)
- DEALing with Image Reconstruction: Deep Attentive Least Squares (ICML 2025 · similarity 0.88)
- A Restoration Network as an Implicit Prior (ICLR 2024 · similarity 0.87)
- pcaGAN: Improving Posterior-Sampling cGANs via Principal Component Regularization (NeurIPS 2024 · similarity 0.86)
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.
