A Novel Composite Network Based on Parallel Dual Attention for Underwater Image Enhancement

Abstract

Due to the absorption and scattering of light by suspended particles, underwater images may suffer from color casts, low contrast, and blurred texture details. Traditional statistics-based and physical model-based methods have improved image quality to some extent, yet they fall short in effectively addressing the complex underwater environment and light conditions. Despite significant improvements in handling complex underwater scenes, existing deep learning-based methods still have limitations in restoring texture details and improving image contrast. To address these issues, a novel composite network is proposed based on parallel dual attention. Firstly, a pair of complementary modules, which consists of a multi-branch color enhancement module and a multi-scale pyramid module, is designed to better extract image features from multiple color channels and multiple scales, respectively. Subsequently, a parallel dual attention module is proposed by combining channel and pixel attention mechanisms to further obtain more useful texture details. Finally, a multi-color space stretch module is used to adaptively increase the contrast of images by adjusting histogram distribution in multiple color spaces. Numerous experiments on public datasets have verified the effectiveness and superiority of our composite network in enhancing different underwater images. Compared with state-of-the-art methods, our method achieves excellent performance on paired datasets in terms of full-reference image quality assessment metrics, and has competitive performance on unpaired datasets as well in terms of reference-free image quality assessment metrics, with minimal computational complexity.