Low-light image enhancement based on normal-light image degradation

In recent years, deep learning has demonstrated its impressive performance in image enhancement. A novel image enhancement method based on normal-light image degradation is proposed in this paper. The degraded images generated from normal-light images by gamma transformation are adopted as the reference images in network training process. Besides, we designed a multi-scale fusion network, which connects two encoding-decoding subnetworks in parallel. The network completes repeated multi-scale fusions by exchanging the information across the parallel subnetworks over and over through the training process. The final enhanced images are obtained by performing inverse gamma transformation on the output of the network. Benefiting from good detail preservation of reference images, smaller gap in brightness and contrast of training image pairs, and the multi-scale fusion network, the method is expected to enhance low-light images while preserving naturalness. Experiments demonstrate the superiority of the proposed method over state-of-the-art image enhancement methods.