Multi-scale error feedback network for low-light image enhancement

作者：Qian, Yi; Jiang, Zetao^*; He, Yuting; Zhang, Shaoqin; Jiang, Shenming

来源：NEURAL COMPUTING & APPLICATIONS, 2022, 34(23): 21301-21317.

DOI：10.1007/s00521-022-07612-8

• Low-light image enhancement
• Error feedback mechanism
• Transformer
• Attention mechanism

Low-light image enhancement is a challenging task because brightness, contrast, noise and other factors must be considered simultaneously. However, most of the existing studies focus on improving illumination, and it is difficult to obtain natural images when the images of complex scenes are enhanced. To address this issue, we propose a neural network-a multi-scale error feedback network (MSEFN)-to enhance low-light images. The proposed network consists of an error feedback encoder module (EFEM), an error feedback decoder module (EFDM) and a feature integration module (FIM). As the main component of EFEM and EFDM, the error feedback feature extraction module can effectively retain spatial information by using the shuffle attention fusion block (SAFB) to fuse the acquired multi-scale features and nonadjacent features. FIM has the ability to capture contextual information that can compensate for the lack of global features in the network. Furthermore, the local uneven illumination (LUI) dataset and polynomial loss function constructed in this paper make our network more stable. Extensive experiments demonstrate that the proposed network outperforms state-of-the-art methods both qualitatively and quantitatively. The LUI dataset is publicly available at: https://github.com/Qyizos/LUI-dataset.

• 单位
  南昌航空大学; 桂林电子科技大学