UDAformer: Underwater image enhancement based on dual attention transformer

• Published in: Computers & graphics Vol. 111; pp. 77 - 88
• Main Authors: Shen, Zhen, Xu, Haiyong, Luo, Ting, Song, Yang, He, Zhouyan
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.04.2023
• Subjects: Self-attention mechanism; Transformer; Underwater image enhancement; Underwater image enhancement; Self-attention mechanism; Transformer
• ISSN: 0097-8493; 1873-7684
• DOI: 10.1016/j.cag.2023.01.009

Underwater images suffer from color casts and low contrast degraded due to wavelength-dependent light scatter and abortion of the underwater environment, which impacts the application of high-level computer vision tasks. Considering the characteristics of uneven degradation and loss of color channel of underwater images, a novel dual attention transformer-based underwater image enhancement method, called UDAformer, is proposed. Specifically, Dual Attention Transformer Block (DATB) combining Channel Self-Attention Transformer (CSAT) with Pixel Self-Attention Transformer is proposed for efficient encoding and decoding of underwater image features. Then, the shifted window method for the pixel self-attention (SW-PSAT) is proposed to improve computational efficiency. Finally, the underwater images are recovered through the design of residual connections based on the underwater imaging model. Experimental results demonstrate the proposed UDAformer surpasses previous state-of-the-art methods, both qualitatively and quantitatively. The code is publicly available at: https://github.com/ShenZhen0502/UDAformer. [Display omitted] •Considering the severe and uneven degradation of underwater images, the SW-PSAT is proposed to model local pixel relationships to restore more image details. Besides, considering the loss of color channels, the CSAT is proposed to extract important features from the different channels of the underwater image.•Considering that a single self-attention mechanism cannot enhance the underwater image well, DATB combined with SW-PSAT and CSAT is proposed to replace the convolution operation of the U-Net architecture.•UDAformer, an encoder–decoder U-shape Transformer for underwater image enhancement, is proposed.