Low-light image enhancement via exploiting original resolution spatial information

As an important class of sensors, cameras can well perceive the real world. However, their outputs, i.e., images, could be significantly degraded by low light. It is of great importance to enhance low-light images. Currently many low-light image enhancement methods are built upon deep learning and often adopt the U-Net structure as backbones. However, the repetitive up-/down-sampling operations of the U-Net structure might destroy the correlation between pixels. To resolve this issue, we propose to exploit the original resolution spatial information in the network design, thus providing better enhancement performance. Specifically, we construct a resolution-preserved and multi-level feature learning framework for low-light image enhancement. On the one hand, the input and output of each basic block have the same resolution as the low-light input, thus preserving the original resolution spatial information for enhancement. On the other hand, each basic block is supported by multi-level feature learning and is therefore able to utilize the contextual information of the UNet structure. Furthermore, we propose a frequency modulation module, which adaptively re-weights the low-/high-frequency components of the features to exploit original resolution spatial information from the frequency perspective. Extensive experiments demonstrate the effectiveness of our proposed enhancement method compared to some state-of-the-art low-light image enhancement methods.