ULD-CycleGAN: An Underwater Light Field and Depth Map-Optimized CycleGAN for Underwater Image Enhancement

Abstract

Underwater imagery frequently exhibits a multitude of degradation phenomena, including chromatic aberrations, optical blurring, and diminished contrast, thereby exacerbating the complexity of underwater endeavors. Among the existing underwater image enhancement (UIE) methods, cycle-consistent generative adversarial network (CycleGAN)-based methods rely on unpaired data sets. Based on CycleGAN, we propose an underwater light field and depth map-optimized CycleGAN (ULD-CycleGAN) for UIE. First, an underwater light field and depth maps are obtained via multiscale Gaussian filtering and the Depth-Net network. Then, they are fed into an enhanced image generator with a dual encoding subnetwork (namely, light-subnet and depth-subnet) for independent encoding. Furthermore, a depth fusion module is designed to enhance the underwater modeling information interaction between these two subnetworks and improve the underwater modeling capabilities of the image enhancement generator. Moreover, a frequency-domain loss is proposed to augment the visual aesthetics of the generated images. Extensive experimental evaluations show that our proposed methodology achieves commendable results in terms of color correction, complex scenes, and luminance, surpassing the state-of-the-art UIE methods in comprehensive qualitative and quantitative assessments. Furthermore, underwater object detection experiments are conducted to further elucidate the efficacy of our ULD-CycleGAN.