Image enhancement in turbid water using multiscale weighted features and attention mechanisms.

In this study, we propose a method to improve the image quality in turbid water by combining multiscale weighted features and attention mechanisms, effectively addressing issues such as reduced contrast and color casts. Initially, global color correction was employed using the grayscale space algorithm. Then, a preprocessed dataset was input into an improved Shallow-UWnet network for feature learning. The model included a multiscale weighted feature fusion module to extract global and local features and a parallel attention module to enhance key details and suppress noise in images. To simulate natural turbid water, a dataset of underwater images with different turbidities was constructed. Our method effectively corrected the color of images in different turbid water conditions and significantly enhanced image quality. In the evaluation on the custom dataset, performance improved by 13%–25% across standard image quality metrics compared with the second-best method. Additionally, we conducted generalization performance tests on the EUVP Dark, UIEBD, and UFO-120 datasets, validating the excellent enhancement effect of our method across different types of images. These results highlight the promising potential of our method for applications in underwater detection, rescue, and marine research.