Research map / Image generation / Super-resolution
ADD: Attribution-Driven Data Augmentation Framework for Boosting Image Super-Resolution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Calibrated Attribution Maps (CAM) for saliency estimation in super-resolution and two saliency-guided data augmentation methods (ADD and ADD+) that paste high-saliency patches across LR-HR pairs to boost SR networks.
Problem
Vanilla data augmentation in low-level vision such as super-resolution causes information loss (flat regions, simple edges), and existing saliency methods (LAM, integrated gradients) are noisy and do not transfer well to SR-based data augmentation.
Approach
A Calibrated Attribution Map combines a Sobel-based global feature detector with calibrated path-integrated gradients over Gaussian-blurred baselines to produce cleaner SR saliency maps; these maps select patches that are cut and mixed across LR-HR pairs (ADD/ADD+) to add diverse degradation patterns.
Key results
- ADD+ improves RealSR PSNR of RCAN by +0.26 dB (29.20 to 29.46) and EDSR by +0.28 dB (28.86 to 29.14) at x4.
- ADD+ improves SwinIR x4 on DIV2K by roughly 0.03 to 0.12 dB across Set5, Set14, Manga109, Urban100 and BSD100.
- Sa2Cor saliency-to-corresponding patch pasting is the best single-patch strategy on DIV2K with EDSR (29.27 dB) and granularity coarsening degrades performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ADD+ during training | SwinIR · Urban100 ×2 | PSNR (dB): 33.29 -> 33.55 (+0.26) | Helped |
| Use ADD+ during training | SwinIR · RealSR ×4 | PSNR: 29.26 -> 29.43 (+0.17) | Helped |
| Use ADD during training | EDSR · LIVE1 q=10 | PSNR: 30.14 -> 30.15 | No clear effect |
| Use ADD during training | EDSR · LIVE1 q=20 | PSNR: 31.83 -> 32.31 | Helped |
| Use ADD during training | EDSR · LIVE1 q=30 | PSNR: 32.45 -> 33.44 | Helped |
| Use 1×1 patch granularity instead of the EDSR baseline (ablation) | EDSR · DIV2K ×4 | PSNR: 29.21 -> 29.28 (+0.07) | Helped |
| Use 2×2 patch granularity instead of the EDSR baseline (ablation) | EDSR · DIV2K ×4 | PSNR: 29.21 -> 29.26 (+0.05) | Helped |
| Use 3×3 patch granularity instead of the EDSR baseline (ablation) | EDSR · DIV2K ×4 | PSNR: 29.21 -> 29.24 (+0.03) | Helped |
| Use 4×4 patch granularity instead of the EDSR baseline (ablation) | EDSR · DIV2K ×4 | PSNR: 29.21 -> 29.22 (+0.01) | No clear effect |
| Use finer patch granularities (5×5–7×7) instead of the EDSR baseline (ablation) | EDSR · DIV2K ×4 | PSNR: 29.14; 29.07; 28.92 | Hurt |
| Add the global feature detector (ablation) | EDSR CAM · DIV2K validation | gradient count at step 50: 120.0k -> 869.1k | Helped |
| Add the global feature detector (ablation) | RCAN CAM · DIV2K validation | gradient count at step 50: 987.3k -> 1291.3k | Helped |
Surprising findings
- Finer patch granularity (5x5 to 7x7) hurts SR performance, with 7x7 falling 0.29 dB below baseline on DIV2K, attributed to boundary effects.
Limitations
- Saliency-based DA transfers poorly from high-level vision without redesign, as discussed in the appendix.
- Gains on standard SR benchmarks are modest (often 0.03 to 0.2 dB).
Experimental setup
Datasets: DIV2K RealSR LIVE1
Benchmarks: Set5 Set14 BSD100 Urban100 Manga109 RealSR DIV2K LIVE1
Metrics: PSNR SSIM insertion/deletion curves diffusion index
Baselines: CutMix CutMixup CutBlur LAM vanilla DA
Builds on: LAM integrated gradients CutBlur CutMix Mixup
Closest papers by meaning
- Interpreting Super-Resolution Networks With Local Attribution Maps (CVPR 2021 · similarity 0.89)
- Rethinking Data Augmentation for Image Super-resolution: A Comprehensive Analysis and a New Strategy (CVPR 2020 · similarity 0.87)
- SaliencyMix: A Saliency Guided Data Augmentation Strategy for Better Regularization (ICLR 2021 · similarity 0.86)
- CAMERAS: Enhanced Resolution and Sanity Preserving Class Activation Mapping for Image Saliency (CVPR 2021 · similarity 0.86)
- Benchmarking the Attribution Quality of Vision Models (NeurIPS 2024 · similarity 0.85)
- BA-SAM: Scalable Bias-Mode Attention Mask for Segment Anything Model (CVPR 2024 · 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.