Research map / Image generation / Image restoration
Color Shift Estimation-and-Correction for Image Enhancement
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 CSEC, a network that enhances images with both over- and under-exposures by estimating and correcting color shifts with a pseudo-normal feature generator, a color-space deformable convolution module (COSE), and a cross-attention color modulation module (COMO).
Problem
Enhancing images that contain both over- and under-exposed regions, where existing methods mishandle color tone shifts because no normal-exposed reference pixels exist.
Approach
Extends deformable convolution to a color space (color offsets per kernel position) to estimate separate color shifts for brightened and darkened features against a generated pseudo-normal feature map, then modulates them with a cross-attention module to produce the output.
Key results
- On LCDP, achieves 23.627 PSNR and 0.855 SSIM, beating the second-best RetinexFormer (23.360 / 0.850) with about 0.30M parameters versus 1.61M.
- On MSEC, reaches 22.728 PSNR and 0.863 SSIM, outperforming FECNet (22.519 / 0.848).
- Achieves the best LAB RMSE on LCDP (6.105), versus 6.148 for RetinexFormer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use color-shift estimation and correction with color modulation | Image enhancement for images with both over- and under-exposures · LCDP test set | PSNR: 23.360 -> 23.627 | Helped |
| use color-shift estimation and correction with color modulation | Image enhancement · MSEC test set | PSNR: 22.519 -> 22.728 | Helped |
| replace color-space deformable convolution with standard convolution (ablation) | COSE in the proposed method · LCDP test set | PSNR: 23.443 -> 23.627 | Helped |
| use spatial deformable convolution with spatial offsets only (ablation) | COSE in the proposed method · LCDP test set | PSNR: 22.623 -> 23.627 | Helped |
| use spatial deformable convolution with spatial offsets and modulation scalars (ablation) | COSE in the proposed method · LCDP test set | PSNR: 23.569 -> 23.627 | Helped |
| use spatial deformable convolution with spatial offsets and color offsets (ablation) | COSE in the proposed method · LCDP test set | PSNR: 23.415 -> 23.627 | Helped |
| replace COMO with the original non-local block (ablation) | The proposed method · LCDP test set | PSNR: 23.201 -> 23.627 | Helped |
| set the over-exposure map to 1 − f(Ix) (ablation) | The proposed method · LCDP test set | PSNR: 21.829 -> 23.627 | Helped |
| use 3-channel illumination maps instead of 1-channel maps (ablation) | The proposed method · LCDP test set | PSNR: 23.016 -> 23.627 | Helped |
| use a pseudo-normal feature generator that does not share parameters with COMO (ablation) | The proposed method · LCDP test set | PSNR: 23.513 -> 23.627 | Helped |
| add pseudo-normal loss (ablation) | The proposed method · LCDP test set | PSNR: 23.588 -> 23.627 | Helped |
| add VGG loss (ablation) | The proposed method · LCDP test set | PSNR: 23.593 -> 23.627 | Helped |
Surprising findings
- Under-exposed and over-exposed pixels show reverse color distribution shifts, motivating separate correction.
- Cosine-similarity regularization used by prior methods is said to fail in large under/over-exposed regions.
Limitations
- Fails when over-exposed pixels are completely saturated, leaving insufficient context to restore colors.
- Relies on the quality of generated pseudo-normal feature maps.
Experimental setup
Datasets: LCDP MSEC MIT-Adobe FiveK Flickr 30K
Benchmarks: LCDP MSEC
Metrics: PSNR SSIM RMSE (LAB) # params
Baselines: HE DSLR HDRNet RetinexNet DeepUPE ZeroDCE MSEC RUAS SNRNet LCDPNet FECNet LANet
Builds on: deformable convolution (DCNv2) non-local block LCDPNet
Closest papers by meaning
- Real-Time Exposure Correction via Collaborative Transformations and Adaptive Sampling (CVPR 2024 · similarity 0.86)
- Region-Aware Exposure Consistency Network for Mixed Exposure Correction (AAAI 2024 · similarity 0.84)
- Color Event Enhanced Single-Exposure HDR Imaging (AAAI 2024 · similarity 0.84)
- ChromaFusionNet (CFNet): Natural Fusion of Fine-Grained Color Editing (AAAI 2024 · similarity 0.84)
- Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.84)
- Bright-NeRF: Brightening Neural Radiance Field with Color Restoration from Low-Light RAW Images (AAAI 2025 · similarity 0.83)
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.