Marine Snow Removal Using a Fully Convolutional 3D Neural Network Combined with an Adaptive Median Filter

Abstract

Marine snow is a type of noise that affects underwater images. It is caused by various biological and mineral particles which stick together and cause backscattering of the incident light. In this paper a method of marine snow removal is proposed. For particle detection a fully convolutional 3D neural network is trained with a manually annotated images. Then, marine snow is removed with an adaptive median filter, guided by the output of the neural network. Experimental results show that the proposed solution is capable of an accurate removal of marine snow without negatively affecting the image quality.