Abstract: The process of obtaining the image with normal exposure time from the image with short
exposure time photographed in extreme low-light conditions is defined as the image enhancement
under extreme low-light conditions. In this paper, we proposed a method for the enhancement of the
extreme low-light image based on the encoding-and-decoding network architecture and residual block.
We designed an end-to-end fully convolutional network as the translation model, which consists of
two parts: the encoding-and-decoding network and refinement network. The input data of the
translation model is the extreme low-light raw data captured with short exposure time in extreme
low-light conditions, and the output data is the image in RGB format. Firstly, the low-frequency
information of the image was reconstructed via U-net and then was input into the residual network to
reconstruct the high-frequency information of the image. Through the experiments carried out on the
SID data set and comparisons with previous research results, it is proved that the method described in
this paper can effectively enhance the visual effect of the images captured under extreme low-light
conditions and improved with low-light enhancement, and increase the expression of the image
details.

Key words: deep learning, convolutional neural network, extremely low-light image, generative
adversarial, image enhancement