Learning-based underwater image enhancement: An efficient two-stream approach

Underwater image enhancement (UIE) attracts extensive attention due to its crucial importance in marine missions. Built upon the powerful representation capacity of deep neural networks (DNNs), recent years have witnessed the exponential growth of learning-based UIE methods. Thus, this work first thoroughly reviews numerous learning-based UIE methods including the algorithm description, dataset generation, and evaluation methodology. Although these learned UIE methods demonstrate remarkable efficiency for improving the reconstruction quality, they still present noticeable limitations, e.g., model generalization and unexpected quality impairments. To tackle these issues, we develop an embarrassingly-efficient two-stream UIE approach to improve both the enhancement quality of underwater images. One stream is the Significant Region Refinement (SRR) that stacks channel attention optimized residual groups to improve the image illumination and resolve the color casting problem, and the other is the Global Appearance Adjustment (GAA) that relies on several dense blocks to enhance the global image sharpness. The final output is derived by intelligently weighing contributions from the SRR and GAA branches. We verify our method using a variety of UIE datasets. Simulations show that our model outperforms state-of-the-art works in both quantitative and qualitative measurements with affordable complexity consumption.

Download Paper