Self-Guided Pixel-Wise Calibration for Low-Light Image Enhancement

Abstract

Unsupervised low-light image enhancement methods have gained attention and shown improvement with low data dependence. However, the lack of a ground truth presents challenges, notably in pronounced noise and color bias. This paper proposes a Self-Guided Pixel-wise Calibration method to overcome associated issues by leveraging inherent features from the input as a self-guide. Specifically, a Pixel-wise Guided Filter is introduced to decrease noise, utilizing a low-light image for guidance and deep features as regularization maps. Additionally, a Color Correction Module is introduced to enhance saturation by adjusting the shadow threshold. Finally, a pixel-wise exposure control loss is formalized to optimize overall naturalness by adjusting brightness to a well-exposedness map from the low-light image. Extensive experiments demonstrate that our method outperforms many state-of-the-art methods, producing enhanced results with fewer distortions across various real-world image enhancement tasks.