LoopExpose: An Unsupervised Framework for Arbitrary-Length Exposure Correction

TL;DR

LoopExpose tackles the challenge of exposure correction without labeled data by introducing a self-reinforcing, two-level nested optimization that uses pseudo-labels from multi-exposure fusion to supervise a luminance-aware exposure correction pathway. A Luminance Ranking Loss leverages sequence luminance ordering to guide learning in a self-supervised manner, while a fixed fusion rule (Mertens) provides stable pseudo-labels for supervision. Empirical results on SeqMSEC and Radiometry512-based benchmarks demonstrate that LoopExpose outperforms existing unsupervised methods and approaches supervised state-of-the-art performance in both single-exposure correction and multi-exposure fusion tasks, with strong qualitative balance across exposure ranges. The framework supports arbitrary-length exposure sequences and can serve as a flexible backbone for SEC and MEF, offering practical scalability by generating training data from RAW batches with fixed exposure offsets. Overall, LoopExpose advances unsupervised exposure correction and fusion, delivering robust performance and paving the way for learnable fusion components in future work.

Abstract

Exposure correction is essential for enhancing image quality under challenging lighting conditions. While supervised learning has achieved significant progress in this area, it relies heavily on large-scale labeled datasets, which are difficult to obtain in practical scenarios. To address this limitation, we propose a pseudo label-based unsupervised method called LoopExpose for arbitrary-length exposure correction. A nested loop optimization strategy is proposed to address the exposure correction problem, where the correction model and pseudo-supervised information are jointly optimized in a two-level framework. Specifically, the upper-level trains a correction model using pseudo-labels generated through multi-exposure fusion at the lower level. A feedback mechanism is introduced where corrected images are fed back into the fusion process to refine the pseudo-labels, creating a self-reinforcing learning loop. Considering the dominant role of luminance calibration in exposure correction, a Luminance Ranking Loss is introduced to leverage the relative luminance ordering across the input sequence as a self-supervised constraint. Extensive experiments on different benchmark datasets demonstrate that LoopExpose achieves superior exposure correction and fusion performance, outperforming existing state-of-the-art unsupervised methods. Code is available at https://github.com/FALALAS/LoopExpose.

Figures (8)

• Figure 1: Illustration of stylistic diversity among expert-retouched reference images in the MSEC dataset msec. The figure also shows the error-exposed inputs with exposure values (EV) of $-1.5$ and $+1.5$, together with the correction result produced by our LoopExpose. While human experts exhibit varying tone and contrast preferences, our method faithfully enhances the original scene, demonstrating strong robustness against subjective editing biases.
• Figure 2: Illustration of the proposed LoopExpose framework. (a) shows the training pipeline based on our nested optimization loop. (b) details the internal components of the SEC model.
• Figure 3: Testing pipeline of LoopExpose. The trained exposure correction model processes arbitrary-length exposure sequences to generate enhanced outputs for fusion or final output.
• Figure 4: Illustration of error accumulation when a learnable MEF network mefnet is used to generate pseudo-labels.
• Figure 5: Visual comparison of exposure correction and fusion results on the MSEC dataset msec. The first row displays the original $+1.5$ EV image and its corresponding corrected results by different methods, while the second row shows the $-1.5$ EV image and its corrected versions. The third row presents the ground truth (GT) and the fused results generated by different MEF algorithms.