Multi-Scale Progressive Fusion Network for Low-Light Image Enhancement

Low-light images affect human perception and vision tasks because of low brightness, loss of details, and severe noise. Most existing methods adopt a multibranch structure with a refusion strategy to solve different image defects separately. However, the correlation between the multi-scale information of images has always been ignored, and the ability of multifeature fusion needs to be improved. In the article, we propose a multi-scale progressive fusion network to obtain feature representation by interacting with different resolution information. Concretely, sampling blocks based on dual-channel superposition are used to acquire different resolution features. We propose a feature fusion block that utilizes local perception and linear correlation to exchange information across resolution layers. An enhancement block based on depth and cyclic residual features is presented to improve brightness and details and suppress noise in different resolution layers. In addition, we introduce a set of loss functions to optimize the model parameters. The proposed method performs better on public datasets and real scenarios.