Region-Aware Exposure Consistency Network for Mixed Exposure Correction

TL;DR

RECNet addresses mixed exposure in single images by learning region-aware representations and bridging exposure gaps between over- and under-exposed regions. It introduces a Region-aware De-exposure Module to map regional features to an exposure-invariant space, a Mixed-scale Restoration Unit to recover details, and an Exposure Contrastive Regularization to enforce intra- and inter-regional consistency. The method achieves state-of-the-art PSNR/SSIM on MSEC and SICE and strong performance on LCDP with mixed exposure, while remaining lightweight. The work provides a practical solution for automatic exposure correction in non-uniform lighting scenes and releases code.

Abstract

Exposure correction aims to enhance images suffering from improper exposure to achieve satisfactory visual effects. Despite recent progress, existing methods generally mitigate either overexposure or underexposure in input images, and they still struggle to handle images with mixed exposure, i.e., one image incorporates both overexposed and underexposed regions. The mixed exposure distribution is non-uniform and leads to varying representation, which makes it challenging to address in a unified process. In this paper, we introduce an effective Region-aware Exposure Correction Network (RECNet) that can handle mixed exposure by adaptively learning and bridging different regional exposure representations. Specifically, to address the challenge posed by mixed exposure disparities, we develop a region-aware de-exposure module that effectively translates regional features of mixed exposure scenarios into an exposure-invariant feature space. Simultaneously, as de-exposure operation inevitably reduces discriminative information, we introduce a mixed-scale restoration unit that integrates exposure-invariant features and unprocessed features to recover local information. To further achieve a uniform exposure distribution in the global image, we propose an exposure contrastive regularization strategy under the constraints of intra-regional exposure consistency and inter-regional exposure continuity. Extensive experiments are conducted on various datasets, and the experimental results demonstrate the superiority and generalization of our proposed method. The code is released at: https://github.com/kravrolens/RECNet.

Figures (9)

• Figure 1: For a single image above in (a), it contains both overexposure (marked with red box) and underexposure (marked with blue box) regions. (b) is the input-ground truth brightness mapping curve statistics of the LCDP wang2022local dataset. (c) and (d) are the generated result-ground truth mapping curves of LCDPNet and our method, respectively. A smaller area represents better results with smaller correction errors.
• Figure 2: The overall architecture of the proposed region-aware exposure correction network, which mainly contains a series of Blocks (RMB) with Region-aware De-exposure Module (RDM) and Mixed-scale Restoration Unit (MRU). The RDM maps exposure features $F_{in}$ to a three-branched exposure-invariant feature $F_{n}$, while the MRU integrates the features $F^{s}$ and $F^{c}$ by the spatial-wise and channel-wise restoration, respectively. The exposure mask predictor (EMP) assists in generating the underexposure feature $F^{u}$ and overexposure feature $F^{o}$. MaIN denotes Mask-aware IN (Instance Normalization) while Refine module represents residual channel attention block. We optimize the model with Exposure Contrastive Regularization (ECR).
• Figure 3: Visualization comparable results on the MSEC dataset of (top) underexposure correction and (bottom) overexposure correction.
• Figure 4: Visualization results on the LCDP dataset of mixed exposure correction. Our model reconstructs the details in the overexposed regions (building and curtain) as well as the underexposed regions (grass and basket).
• Figure 5: The visualization of the input/output feature of RMB module.