Research map / Image generation / Image restoration
QuARF: Quality-Adaptive Receptive Fields for Degraded Image Perception
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 QUARF, a plug-and-play multi-kernel convolutional block whose kernel routing is conditioned on blind image quality features, to make DNNs robust to degraded inputs without hurting pristine performance.
Problem
DNN performance drops sharply on degraded images, and data augmentation to fix this tends to hurt performance on pristine images.
Approach
Multiscale continuous receptive fields (dilated 3x3 convolutions emulating kernel sizes 3,5,9,17) are grouped by channel split, and a quality-adaptive router uses BIQA features (CONTRIQUE) to softmax-weight the kernel outputs per group, so the inference route adapts to input quality.
Key results
- On CelebAMask-HQ face parsing with SegNeXt, QUARF raises average mIoU from 59.77 (data aug) to 60.03 and average BIoU from 47.70 to 48.62 over degradation levels.
- On MATEBIT exemplar-based translation, FID drops from 35.14 (data aug) to 29.87 with QUARF; on Pix2Pix FSS, FID goes from 40.54 (original) to 26.25.
- On Real-ESRGAN SR, PSNR on Urban100 rises from 24.03 to 26.67 with QUARF versus the official model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add QUARF instead of data augmentation | Pix2Pix · Face sketch synthesis | FID: 26.68 -> 26.25 | Helped |
| Add QUARF instead of data augmentation | CycleGAN · Summer-to-winter translation | FID: 80.41 -> 83.97 | Hurt |
| Add QUARF instead of data augmentation | GLD · Image-to-sketch translation | SIFID: 1.763 -> 1.245 | Helped |
| Add QUARF instead of data augmentation | MATEBIT · Exemplar-based translation | FID: 35.14 -> 29.87 | Helped |
| Add QUARF instead of data augmentation | AdaAttN · Neural style transfer | FID: 89.39 -> 86.39 | Hurt |
| Add QUARF to Real-ESRGAN | Real-ESRGAN · BSD100 ×4 super-resolution | PSNR: 25.21 -> 26.69 | Helped |
| Add QUARF to AST | AST · AGAN-Data deraining | PSNR: 32.32 -> 32.39 | Helped |
| Add QUARF instead of data augmentation | SegNeXt · CelebAMask-HQ average LQ | mIoU: 59.77 -> 60.03 | Helped |
| Place QUARF in the encoder rather than the decoder (ablation) | Pix2Pix · Face sketch synthesis | FID: 26.89 -> 26.25 | Helped |
| Use CONTRIQUE quality features rather than CNN quality features (ablation) | QUARF on CycleGAN · Summer-to-winter translation | FID: 133.64 -> 83.97 | Helped |
| Use CONTRIQUE quality features rather than BRISQUE quality features (ablation) | QUARF on MATEBIT · Exemplar-based translation | SIFID: 0.610 -> 0.437 | Helped |
| Use 2 to 8 channel groups rather than other tested group counts (ablation) | QUARF on MATEBIT · MATEBIT | FID | Mixed |
Surprising findings
- Data augmentation reduces performance on pristine images for several models (e.g., HQ mIoU drops from 74.51 to 71.79 for SegNeXt).
Limitations
- Gains on some metrics are mixed; SIFID and CPBD are not uniformly best, and QUARF hurt HQ mIoU for HQ-SAM (88.67 to 87.70).
- Degradation evaluated mainly with hybrid/Gaussian blur synthetic corruptions, not real-world degradations.
Experimental setup
Datasets: CUFS summer2winter Metface COCO Anime Colorization MSCOCO WikiArt DF2K OST300 AGAN-Data CelebAMask-HQ HQSeg-44K
Benchmarks: CUFS summer2winter Metface AGAN-Data DF1K BSD100 Manga109 Urban100 CelebAMask-HQ ThinObject-5K DIS HR-SOD
Metrics: FID SIFID CPBD PSNR SSIM LPIPS mIoU BIoU
Baselines: Official models Data Augmentation SKNet LSKNet Inception OCNet MANet BRISQUE-based QUARF variant
Builds on: SegNeXt Real-ESRGAN HQ-SAM CONTRIQUE dilated convolution
Closest papers by meaning
- Receptive Fields As Experts in Convolutional Neural Architectures (ICML 2024 · similarity 0.87)
- Q-Insight: Understanding Image Quality via Visual Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- Q-Probe: Scaling Image Quality Assessment to High Resolution via Context-Aware Agentic Probing (NeurIPS 2026 · similarity 0.85)
- UARE: A Unified Vision-Language Model for Image Quality Assessment, Restoration, and Enhancement (CVPR 2026 · similarity 0.85)
- Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment (CVPR 2025 · similarity 0.84)
- Q-Bench: A Benchmark for General-Purpose Foundation Models on Low-level Vision (ICLR 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.