Research map / Image generation / Autoregressive and latent generation
Rethinking No-reference Image Exposure Assessment from Holism to Pixel: Models, Datasets and Benchmarks
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 P-IEANet, a Haar-DWT-based pixel-level no-reference image exposure assessment network, together with the IEA40K dataset (40K images with pixel-level expert labels) and a 19-method benchmark.
Problem
No-reference image exposure assessment is inaccurate, poorly generalizable, and limited to holistic scores, lacking fine-grained pixel-level evaluation of complex exposure conditions.
Approach
Decomposes images with Haar DWT into low-frequency lightness and high-frequency structure features, then predicts a per-pixel exposure residual relative to an implicit ideal exposure, trained on pixel-level expert annotations.
Key results
- P-IEANet reaches MAE 0.03 and SSIM 0.75 on IEA40K pixel-level assessment, versus 0.05 and 0.60 for the best baseline (LIT), with about 97% fewer parameters.
- On SPAQ holistic assessment, P-IEANet achieves LRCC 0.78 and SRCC 0.73, above Q-align (0.75 / 0.70).
- Adding P-IEANet as a frozen evaluator raises Retinex and GSAD on LOL-v1 (e.g. GSAD* PSNR 27.1 to 28.2 with P-IEANet).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Haar DWT (ablation) | P-IEANet · IEA40K | SSIM: 0.43 -> 0.75 | Helped |
| Add the Structure Feature Module (ablation) | P-IEANet · IEA40K | SSIM: 0.54 -> 0.75 | Helped |
| Add the Lightness Feature Module (ablation) | P-IEANet · IEA40K | SSIM: 0.42 -> 0.75 | Helped |
| Replace Haar wavelet with Daubechies (ablation) | P-IEANet · IEA40K | SSIM: 0.49 vs 0.75 | Hurt |
| Replace Haar wavelet with Symlet (ablation) | P-IEANet · IEA40K | SSIM: 0.41 vs 0.75 | Hurt |
| Replace L1-norm loss with L2-norm (ablation) | P-IEANet · IEA40K | SSIM: 0.62 vs 0.75 | Hurt |
| Replace L1-norm loss with Smooth L1 (ablation) | P-IEANet · IEA40K | SSIM: 0.67 vs 0.75 | Hurt |
| Add P-IEANet exposure-residual loss | Retinex* · LOL-v1 | PSNR: 24.7 -> 25.3 | Helped |
| Add P-IEANet exposure-residual loss | Retinex* · LOLv2-real | SSIM: 0.82 -> 0.86 | Helped |
| Add P-IEANet exposure-residual loss | GASD* · LOLv2-real | PSNR: 28.4 -> 29.5 | Helped |
| Use P-IEANet for pixel-level assessment instead of LIT | IEA40K pixel-level assessment · IEA40K | MAE: 0.05 -> 0.03 | Helped |
| Use P-IEANet for holistic scoring instead of Q-align | SPAQ holistic assessment · SPAQ | SRCC: 0.71 -> 0.73 | Helped |
Surprising findings
- Without any fine-tuning, the pixel residual averaged into a holistic score reaches LRCC 0.69 and SRCC 0.65 on SPAQ, exceeding some fine-tuned baselines.
Limitations
- Few images of high-speed moving objects, so misalignment is hard to handle; such cases are simulated by EV adjustment.
- Resizing high-resolution images to a 512-pixel short side may degrade pixel-level assessment.
- Holistic results are partly reported with LRCC/SRCC only; paper states the dataset includes simulated exposures for 80% of images.
Experimental setup
Datasets: IEA40K SPAQ LOL-v1 LOLv2-real MIT-Adobe FiveK
Benchmarks: IEA40K SPAQ LOL-v1 LOLv2-real
Metrics: SSIM MAE PSNR LCC SRCC ACC
Baselines: FEC MSLT Retinex PyDiff SMG SKF GSAD LIT ArnIQA ReIQA DEIQT MUSIQ
Builds on: Haar discrete wavelet transform Adams zone system
Closest papers by meaning
- Toward Generalized Image Quality Assessment: Relaxing the Perfect Reference Quality Assumption (CVPR 2025 · similarity 0.86)
- Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · similarity 0.83)
- Bricker to BRACE: A Bracket Exposure RAW Dataset and Restoration Model for Flicker-Banding (NeurIPS 2026 · similarity 0.82)
- RIFLE: Removal of Image Flicker-Banding via Latent Diffusion Enhancement (NeurIPS 2026 · similarity 0.82)
- Real-Time Exposure Correction via Collaborative Transformations and Adaptive Sampling (CVPR 2024 · similarity 0.82)
- Region-Aware Exposure Consistency Network for Mixed Exposure Correction (AAAI 2024 · similarity 0.82)
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.