Underwater Image Enhancement by Combining Multi-Attention with Recurrent Residual Convolutional U-Net

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.