Abstract:
Poor lighting conditions result in photographs with low contrast. Most existing methods focus on low-light enhancement and perceptual quality improvement, and have not taken both under- and overexposure into consideration. In this paper, we propose a novel semi-supervised learning method for single image contrast enhancement. The supervised branch is trained using paired data under the constraint of supervised losses. While in the unsupervised branch, we explore content consistency and illumination prior as loss functions to train the network. The advantages of the proposed approach are two folds. First, guided by ground truth images, the supervised branch learns well to preserve image details and suppress noise. Second, the unsupervised branch learns to adapt to more illumination intensities and diverse illumination environments, which bridges the gap between various lighting conditions. With the help of the semi-supervised strategy, our method uses a single model to enhance both underexposed and overexposed images, and generalizes well to various lighting conditions. Experimental results show that the proposed method outperforms the state-of-the-art methods quantitatively and qualitatively.
