LAU-Net: A low light image enhancer with attention and resizing mechanisms


Lim, Choon Chen and Loh, Yuen Peng and Wong, Lai Kuan (2023) LAU-Net: A low light image enhancer with attention and resizing mechanisms. Signal Processing: Image Communication, 115. p. 116971. ISSN 0923-5965

[img] Text
4.pdf - Published Version
Restricted to Repository staff only

Download (3MB)


Nighttime environments with sub-optimal lighting conditions significantly degrade the quality of captured images. Even though many notable state-of-the-art methods had been proposed to enhance low-light images, many of the enhanced outcomes exhibit color distortion, and uneven light adjustment problems. To remedy these issues, we propose an effective supervised network, Low-light Advanced U-Net (LAU-Net), which restructures the regular U-Net to offer a better network for low-light image enhancement. Specifically, we merged several efficacious components into our LAU-Net, namely the Parallel Attention Unit (PAU), the Internal Resizing Module (IRM), and external convolutional layers. The PAU places two attention modules in parallel to extract features along the convolutional streams. Meanwhile, the IRM comprises resizing components to optimize the information flow from encoder blocks to decoder blocks, whereas the external convolutional layers simulate the autoencoder to suppress noises. We employed the LOL dataset, which is composed of 500 paired images, to train, validate, and test the proposed network. Rigorous experiments showed that our model delivered remarkable performance both in qualitative and quantitative assessments and outperforms state-ofthe-art approaches. Moreover, ablation studies also justified the necessity of each module in our proposed design. Lastly, we demonstrated that the proposed method could serve as an excellent pre-processing tool for image classification tasks in challenging nighttime environments, as it has successfully improved the object classification accuracy of a ResNet-50 model when applied onto low-light images from the ExDark dataset

Item Type: Article
Uncontrolled Keywords: Low-light image enhancement Advanced U-Net Attention Resizing modules
Subjects: G Geography. Anthropology. Recreation > GE Environmental Sciences (General) > GE1-350 Environmental sciences > GE300-350 Environmental management
Divisions: Faculty of Computing and Informatics (FCI)
Depositing User: Ms Nurul Iqtiani Ahmad
Date Deposited: 02 Jun 2023 00:45
Last Modified: 02 Jun 2023 00:45


Downloads per month over past year

View ItemEdit (login required)