Abstract

Water and light interactions cause color shifts and blurring in underwater images, while dynamic underwater illumination further disrupts scene consistency, resulting in poor performance of optical image-based reconstruction methods underwater. Although Neural Radiance Fields (NeRF) can describe aqueous medium through volume rendering, applying it directly underwater may induce artifacts and floaters. We propose SP-SeaNeRF, which uses micro MLP to predict water column parameters and simulates the degradation process as a combination of real colors and scattered colors in underwater images, enhancing the model's perception of scattering. We use illumination embedding vectors to learn the illumination bias within the images, in order to prevent dynamic illumination from disrupting scene consistency. We have introduced a novel sampling module, which focuses on maximum weight points, effectively improves training and inference speed. We evaluated our proposed method on SeaThru-NeRF and Neuralsea underwater datasets. The experimental results show that our method exhibits superior underwater color restoration ability, outperforming existing underwater NeRF in terms of reconstruction quality and speed.