Deep Symmetric Network for Underexposed Image Enhancement

with Recurrent Attentional Learning

Lin Zhao1*, Shao-Ping Lu1*#, Tao Chen2, Zhenglu Yang1, Ariel Shamir2

1TKLNDST, CS, Nankai University, Tianjin, China

2Elephant Technologies, China

3The Interdisciplinary Center, Herzliya, Israel;;;






Underexposed image enhancement is of importance in many research domains. In this paper, we take this problem as image feature transformation between the underexposed image and its paired enhanced version, and we propose a deep symmetric network for the issue. Our symmetric network adapts invertible neural networks (INN) for bidirectional feature learning between images, and to ensure the mutual propagation invertible we specifically construct two pairs of encoder-decoder with the same pre-trained parameters. This invertible mechanism with bidirectional feature transformations enable us to both avoid colour bias and recover the content effectively for image enhancement. In addition, we propose a new recurrent residual-attention module (RRAM), where the recurrent learning network is designed to gradually perform the desired colour adjustments. Ablation experiments are executed to show the role of each component of our new architecture. We conduct a large number of experiments on two datasets to demonstrate that our method achieves the state-of-the-art effect in underexposed image enhancement. Code is available at


The results of different methods on challenging images:











The visual results of different methods for the underexposed image:















Comparison results with image retouching methods:












The results of low-light image enhancement methods:










Visual results of different ablation experiments on our method:







Quantitative results of different methods:













License. It is open source under BSD-3 license. Codes can be used freely only for academic or education purpose. If you want to apply it to industrial products, please send an email to Shao-Ping Lu at first.


In the camera-ready of our ICCV paper, there are mistakes for the DeepLPF results in Fig. 1(c), Fig. 5(n), Fig. 6(g) and Tab. 1. The reason is that we trained 300 epochs with the batch size as 30 (we randomly cropped each image to 250x250 to set multiple batch sizes for training), while the original set of the DeepLPF code is batch_size = 1. After the DeepLPF authors reminded us, now we set the batch size to 1 and re-trained for 300 epochs to correct the results of the figures and tables in the paper.


Deep Symmetric Network for Underexposed Image Enhancement with Recurrent Attentional Learning, Lin Zhao*, Shao-Ping Lu*#, Tao Chen, Zhenglu Yang, Ariel Shamir, ICCV 2021. [PDF] [Chinese]