Source-linked AI summary
Dual Illumination Estimation for Robust Exposure Correction
Qing Zhang, Yongwei Nie, Wei-Shi Zheng
TL;DR
Exposure correction must handle varied lighting conditions, yet existing methods may be visually unsatisfactory or limited to particular exposure types. This paper estimates illumination for both an image and its inversion, then fuses the resulting corrections with the input. The authors report effective, high-quality correction across challenging exposure conditions and favorable comparisons with existing tools and methods.
Problem
Existing exposure-correction approaches can produce unsatisfactory results, target limited exposure types, or depend on training images unlike the test images.
Method
The method performs dual illumination estimation on the input and inverted input, recovers complementary corrections, and fuses them with the original image.
Results
The method produces high-quality correction for challenging images, is preferred by human subjects, and performs comparably to expert retouching in the reported user study.
Takeaways & Limitations
The approach provides a fully automatic, near-interactive exposure-correction tool for images containing underexposed, overexposed, or mixed regions.
Takeaways & Limitations
The method may amplify noise together with fine-scale details when the input image is noisy.
Abstract
from arXiv · showhide
Exposure correction is one of the fundamental tasks in image processing and computational photography. While various methods have been proposed, they either fail to produce visually pleasing results, or only work well for limited types of image (e.g., underexposed images). In this paper, we present a novel automatic exposure correction method, which is able to robustly produce high-quality results for images of various exposure conditions (e.g., underexposed, overexposed, and partially under- and over-exposed). At the core of our approach is the proposed dual illumination estimation, where we separately cast the under- and over-exposure correction as trivial illumination estimation of the input image and the inverted input image. By performing dual illumination estimation, we obtain two intermediate exposure correction results for the input image, with one fixes the underexposed regions and the other one restores the overexposed regions. A multi-exposure image fusion technique is then employed to adaptively blend the visually best exposed parts in the two intermediate exposure correction images and the input image into a globally well-exposed image. Experiments on a number of challenging images demonstrate the effectiveness of the proposed approach and its superiority over the state-of-the-art methods and popular automatic exposure correction tools.
1. Introduction
Exposure correction remains difficult because poorly exposed images are visually unpleasing, existing controls are tedious for non-experts, and prior methods often target only under- or over-exposure. The paper introduces dual illumination estimation and multi-exposure fusion to automatically correct diverse exposure conditions.
- Poorly exposed photos often have unclear details, weak contrast, and dull color, motivating effective exposure correction.
- Interactive editing tools require non-experts to balance multiple controls, while one-click Auto-Tone and Auto-Level may produce unsatisfactory adjustments.
- Many existing methods focus solely on under- or over-exposure, limiting their applicability to images with varied exposure conditions.
- Dual illumination estimation treats under- and over-exposure correction as illumination estimation on the input and inverted input images.
- The method recovers two intermediate corrections and fuses them with the input image to blend locally best-exposed regions into a globally well-exposed result.
- Experiments report stronger human preference, effectiveness on images containing both under- and over-exposed regions, full automation, and near-interactive runtime.
2. Related Work
Prior exposure-correction approaches include histogram manipulation, exposure fusion, tone mapping, interactive and semantic methods, but they face constraints such as unrealistic results, input requirements, or sensitivity to extracted semantics. The proposed framework addresses overexposure through inverted-image illumination estimation and combines corrected exposures through fusion.
- Histogram equalization improves contrast by stretching intensity distributions but can generate unrealistic results because it ignores relationships between pixels.
- Exposure fusion can blend well-exposed regions, but classic methods require a bracketed multi-exposure sequence rather than a single image.
- The proposed pipeline estimates forward and reverse illuminations, recovers two corrected images, and fuses them with the input image.
- Layer-based and region-based tone-mapping methods can preserve details or target local exposure, yet may produce oversaturation or depend on difficult region segmentation.
- Reverse tone mapping infers HDR from an LDR image, whereas this work preserves the input bit depth and produces results directly displayable on devices.
- Interactive correction methods benefit from user input, while the proposed approach is fully automatic.
- Content-aware methods use image semantics for areas such as faces, skin, and sky, but their quality is sensitive to semantic extraction reliability.
- Learning-based enhancement methods may perform poorly on images that differ substantially from their training images.
3. Our Approach
The approach performs dual illumination estimation to correct under- and over-exposed regions separately, then fuses the input and corrected images using exposure-aware quality maps. Its refined illumination preserves salient structures while suppressing redundant texture details, and modified fusion maps improve local detail and color.
- Dual illumination estimation: Retinex-based recovery represents an image as the pixel-wise product of an enhanced image and a single-channel illumination map.Once illumination is estimated, the enhanced image is recovered by pixel-wise division through the illumination map.
- Dual illumination estimation: Dual illumination estimation applies forward estimation to the input and reverse estimation to its inversion, targeting under- and over-exposed regions respectively.Inversion makes originally overexposed regions appear underexposed, enabling their correction through illumination estimation.
- Illumination estimation framework: The refined illumination objective stays close to the initial illumination while minimizing partial derivatives to remove redundant texture details.Its smoothness weights are designed differently from WLS and are inspired by relative total variation.
- Illumination estimation framework: Figure 7 compares WLS, RTV, and the proposed illumination estimation from the same initial illumination, with Gamma adjustment applied to recovered results.The proposed estimate better removes redundant texture while preserving salient illumination structures.
- Multi-exposure image fusion: Multi-resolution fusion blends the input with forward- and reverse-corrected images to retain locally best-exposed parts across the image sequence.Replacing normalized quality maps with per-pixel maxima improves face and cloud details, contrast, and color vividness.
4. Experiments
Two user studies compare the method with automatic tools, expert retouching, and learning-based methods on challenging exposure conditions. Results indicate stronger human preference and visually appealing corrections, while analysis also identifies failure cases and noise amplification.
- User study #1: The first user study compared the method with Photoshop Auto-Level, Lightroom Auto-Tone, interactive Lightroom correction, and expert retouching using 100 MIT-Adobe FiveK images rated by 100 participants.Image pairs were presented anonymously and in random order to reduce subjective bias.
- User study #1: Human subjects preferred the method over the compared automatic tools, with preference comparable to expert retouching.The authors interpret this as evidence that the method can produce high-quality corrections for non-expert users.
- User study #1: Visual examples span overexposed, underexposed, and mixed-exposure scenes, where the method produced normal brightness, clear details, distinct contrast, and vivid color comparable to expert-retouched results.The examples include overexposed apples, an underexposed building with an overexposed sky, an overexposed wedding image, and a globally underexposed portrait.
- User study #2: The second user study compared the method with HDRNet, DPE, and Exposure on 100 test images selected for substantial underexposed or overexposed regions.The comparisons used publicly available implementations and pairwise preference ratings from 100 participants.
- User study #2: More participants favored the method than the compared learning-based methods, while visual examples show competing methods failing on some overexposed or unrealistic results.HDRNet sometimes increased exposure in overexposed portraits, DPE could generate unrealistic results, and Exposure struggled with overexposed images.
- More analysis: The method differs from intrinsic image decomposition because its recovered I′ is intended as a natural-looking exposure correction rather than illumination-independent reflectance.A comparison with an IID reflectance result is shown using an IIW image.
- More analysis: The authors report failures when facial regions are nearly black or white without recoverable color and texture, and note that noise may be amplified with fine-scale details.Other state-of-the-art methods also fail on the cited facial examples.
5. Conclusion and Future Work
The method combines dual illumination estimation with multi-exposure fusion to produce globally well-exposed images automatically. The authors identify noise suppression and recovery of missing content in extremely exposed regions as future directions.
- Conclusion: Dual illumination estimation produces intermediate corrections for under- and over-exposed regions, which multi-exposure fusion integrates with the input image.The fusion selects locally best-exposed parts from all three images to form a globally well-exposed result.
- Conclusion: The method is simple, fully automatic, and operates at a near-interactive rate.
- Conclusion: Extensive comparisons with automatic exposure tools and state-of-the-art methods are used to demonstrate the method’s effectiveness.
- Future Work: The authors plan to suppress noise during exposure correction.
- Future Work: Recovering missing content in extremely under- and over-exposed regions remains an identified future direction.The paper illustrates such failures in facial regions and proposes semantics information and texture synthesis as possible tools.