Removal of Floating Particles from Underwater Images Using Image Transformation Networks

In this paper, we propose three methods for removing floating particles from underwater images. The first two methods are based on Generative Adversarial Networks (GANs). The first method uses CycleGAN which can be trained with an unpaired dataset, and the second method uses pix2pixHD that is trained with a paired dataset created by adding artificial particles to underwater images. The third method consists of two-step process – particle detection and image inpainting. For particle detection, an image segmentation neural network U-Net is trained by using underwater images added with artificial particles. Using the output of U-Net, the particle regions are repaired by an image inpainting network Partial Convolutions. The experimental results showed that the methods using GANs were able to remove floating particles, but the resolution became lower than that of the original images. On the other hand, the results of the method using U-Net and Partial Convolutions showed that it is capable of accurate detection and removal of floating particles without loss of resolution.