Learning Multi-scale Retinex with Residual Network for Low-Light Image Enhancement

Long Ma,Jie Lin,Jingjie Shang,Wei Zhong,Xin Fan,Zhongxuan Luo,Risheng Liu
DOI: https://doi.org/10.1007/978-3-030-60633-6_24
2020-01-01
Abstract:Existing network-based techniques addressing low-light image enhancement are developed by using the complex network architectures. However, the performance is not ideal, and they cannot give a reasonable interpretation of the effects of each layer in the network. To settle these issues, we design a residual network to learn multi-scale Retinex for handling low-light image enhancement. To be concrete, inspired by multi-scale Retinex, we define a new residual-type multi-scale Retinex model to gradually remove the illumination generated by the convolutional procedure. Thanks to the progressive mechanism, we can build an intuitive and explicit relationship between our residual-type multi-scale Retinex model and residual network. This enables us can directly utilize the residual network to learn a residual-type multi-scale Retinex by integrating the data distribution. Precisely because of our transparent modeling procedure, we can recognize the effects of each layer in our learnable architecture. It is valuable for more effectively exploit the network layers to handle this task. Extensive analytical experiments are performed to verify the effectiveness of our proposed method. A series of evaluative experiments are conducted to illustrate our superiority against other state-of-the-art methods.
What problem does this paper attempt to address?