A Semi-supervised Low-Light Image Enhancement with Color Guidance

Low-light image enhancement (LLIE) has found extensive applications in surveillance and related fields. Current deep learning-based algorithms have demonstrated significant advantages over traditional methods for LLIE. However, the issue of color distortion still remains a challenge. This paper proposes a semi-supervised curve estimation and restoration network that consists of two cascaded sub-networks with distinct functionalities. The first-level sub-network applies specific constraints to enhance different regions based on their brightness levels. The second-level sub-network is trained using superpixel color consistency loss and global histogram loss to effectively correct color distortion. Experimental results demonstrate that the proposed method achieves satisfactory outcomes in both visual quality and evaluation metrics.