Learning to Enhance Low-Light Images with a Synthetic-Real Interaction Training Strategy

摘要

Recently, utilizing deep networks to improve the visual quality of low-light images has attracted widespread attention. These works attach importance to design the network architecture and loss functions. However, they heavily overlook the strong portrayal ability of the network and a critical factor that decides the performance of deep networks, i.e., training strategy. In this work, we design a kind of effective training strategy, rather than meticulously designing the architectures and loss functions. Concretely, we define a lightweight illumination estimation network which cannot obtain the ideal performance by directly using the real paired low-lightimages. We design a synthetic-real interaction training strategy consists of synthetic and real training phases. We synthesize the low-light image pairs by treating the depth map as the illumination. In this way, we can provide the supervised labels for simulating the smoothed property of the illumination, avoid designing the complex regularization loss which may be minor effects. Next, we fix the trained network parameters to further train it by utilizing the real paired low-light images. Actually, we can repeatedly and alternatively execute synthetic-real interaction training mechanisms to obtain more ideal performance. Extensive analyses and evaluations fully verify our effectiveness and outstanding outcomes against other state-of-the-art deep networks.