LoopExpose: An Unsupervised Framework for Arbitrary-Length Exposure Correction

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Supervised learning for exposure correction requires extensive labeled datasets, which are often impractical to acquire.
• Challenging lighting conditions significantly degrade image quality, necessitating effective correction methods.

Purpose of the Study:

• To propose LoopExpose, a novel pseudo label-based unsupervised method for arbitrary-length exposure correction.
• To address the limitations of supervised methods by eliminating the need for labeled data.

Main Methods:

• A nested loop optimization strategy jointly optimizes the correction model and pseudo-supervised information.
• A feedback mechanism refines pseudo-labels through a self-reinforcing learning loop.
• A Luminance Ranking Loss is introduced as a self-supervised constraint leveraging relative luminance.

Main Results:

• LoopExpose demonstrates superior performance in exposure correction and image fusion compared to existing unsupervised methods.
• Experiments on benchmark datasets validate the effectiveness of the proposed approach.
• The method achieves state-of-the-art results in unsupervised exposure correction.

Conclusions:

• LoopExpose offers an effective unsupervised solution for exposure correction, particularly valuable when labeled data is unavailable.
• The proposed nested loop optimization and Luminance Ranking Loss contribute to robust performance.
• The method advances the field of image quality enhancement under adverse lighting conditions.