Underwater image enhancement by combining multi-attention with recurrent residual convolutional U-Net

Shuqi Wang,Zhixiang Chen,Hui Wang

DOI: https://doi.org/10.1007/s11760-023-02985-2

2024-01-30

Abstract:The scattering and absorption of light lead to color distortion and blurred details in the captured underwater images. Although underwater image enhancement algorithms have made significant breakthroughs in recent years, enhancing the effectiveness and robustness of underwater degraded images is still a challenging task. To improve the quality of underwater images, we propose a combined multi-attention mechanism and recurrent residual convolutional U-Net (ACU-Net) for underwater image enhancement. First, we add a dual-attention mechanism and convolution module to the U-Net encoder. It can unequally extract features in different channels and spaces and make the extracted image feature information more accurate. Second, we add an attention gate module and recurrent residual convolution module to the U-Net decoder. It helps extract features fully and facilitates the recovery of more detailed information when the image is generated. Finally, we test the subjective results and objective evaluation of our proposed algorithm on synthetic and real datasets. The experimental results show that the robustness of our algorithm outperforms the other five classical algorithms, such as in enhancing underwater images with different color shifts and turbidity. Moreover, it corrects the color bias and improves the contrast and detailed texture of the images.

engineering, electrical & electronic,imaging science & photographic technology

What problem does this paper attempt to address?

The paper aims to address the issues of color distortion and detail blurring in underwater images caused by light scattering and absorption. Despite significant progress in underwater image enhancement algorithms in recent years, challenges remain in improving image quality and robustness. To tackle these problems, the authors propose an underwater image enhancement method that combines multiple attention mechanisms with a cyclic residual convolutional U-Net (ACU-Net).

Specifically, the main contributions of this method include:

1. **Improved Feature Extraction**: Using U-Net, which has excellent feature extraction capabilities, as the base network, and introducing dual attention mechanisms and convolution modules in the encoder to handle features in different channels and spaces unequally, thereby more accurately extracting image feature information.

2. **Application of Attention Mechanisms**: Adding attention gate modules and cyclic residual convolution modules in the U-Net decoder to help fully extract features and promote the recovery of more detailed information during the generation process.

3. **Experimental Validation**: Conducting subjective and objective evaluations on synthetic and real datasets, the results show that the proposed algorithm outperforms five other classic algorithms in enhancing underwater images with different color shifts and turbidity, effectively correcting color deviations, improving contrast, and detailed textures.

In summary, this study proposes a new underwater image enhancement strategy—combining multiple attention mechanisms with a cyclic residual convolutional U-Net method—to improve the quality of low-quality underwater images, focusing on extracting sufficient features without generating excessive noise. This method not only effectively improves the visual effect of images but also maintains good performance under different levels of water pollution and color shifts.