Temporally Averaged Regression for Semi-Supervised Low-Light Image Enhancement.

Constructing annotated paired datasets for low-light image enhancement is complex and time-consuming, and existing deep learning models often generate noisy outputs or misinterpret shadows. To effectively learn intricate relationships between features in image space with limited labels, we introduce a deep learning model with a backbone structure that incorporates both spatial and layer-wise dependencies. The proposed model features a baseline image-enhancing network with spatial dependencies and an optimized layer attention mechanism to learn feature sparsity and importance. We present a progressive supervised loss function for improvement. Furthermore, we propose a novel Multi-Consistency Regularization (MCR) loss and integrate it within a Multi-Consistency Mean Teacher (MCMT) framework, which enforces agreement on high-level features and incorporates intermediate features for better understanding of the entire image. By combining the MCR loss with the progressive supervised loss, student network parameters can be updated in a single step. Our approach achieves significant performance improvements using fewer labeled data and unlabeled low-light images within our semi-supervised framework. Qualitative evaluations demonstrate the effectiveness of our method in leveraging comprehensive dependencies and unlabeled data for low-light image enhancement.