Physically-based Two-stage Underwater Image Restoration Nan Wang 1 , Lin Qi 1 , Junyu Dong 1* , Hao Fan 1 , Xingnan Chen 1 , Hui Yu 2 1 Ocean University of China 2 University of Portsmouth dongjunyu@ouc.edu.cn ABSTRACT Underwater images are blurred due to light scattering and absorption. Image restoration is therefore important in many underwater research and practical tasks. In this paper, we propose an effective two-stage method to restore underwater scene images. Based on an underwater light propagation model, we first remove backscatter by fitting a binary quadratic function. Then we eliminate the forward scattering and non-uniform lighting attenuation using blue-green dark channel prior. The proposed method requires no additional calibration and we show its effectiveness and robustness by restoring images captured under various underwater scenes. Keywords: Underwater image restoration; backscatter removal; dark channel prior; underwater lighting propagation model. 1. INTRODUCTION Light propagation in the water is heavily attenuated due to the absorption and scattering by numerous suspending particles. Underwater images are accordingly blurred. Restoration of underwater images is therefore important for oceanographic exploration and computer vision tasks such as underwater 3D reconstruction. The key to underwater image restoration is to remove the influence of absorption and scattering. In addition to studying light transmission in the water [1, 2, 3, 4], researchers also made efforts in investigating how light propagates in different medium environment, like hazy and foggy scenes [5, 6, 7]. A challenging problem of underwater image restoration is backscatter removal. Some work proceeds with directly estimating the parameters in traditional optical propagation model [1, 4]. And other approaches indirectly solve the complex problem with the assist of polarization [8, 9] or fitting some specific function like the binary quadratic function [3], achieving good result as well. Traditional dehazing techniques have been applied in underwater scenes as their propagation models share some similarities. Dark channel prior [7], which has been widely researched in many fields, was also found to be able to improve the quality of underwater images [10]. Different methods mentioned above appear with distinctive features, but also limitations exist. The estimation of complex optical propagation parameters calls for specific scenes assumption and necessary approximation, which brings strict limit with complicated calculation. Dark channel prior method has a certain effect in underwater images processing [10], but the underwater propagation model is different from the dehazing model, which contains dominant component of backscatter. In this paper, we propose a novel two-stage approach to restore underwater images without additional pre-requisite calibration. First, we improve the backscatter removal method in [3] by separating the background pixels from the image, from which we use a more reasonable function fitting method to estimate the backscatter. Second, we propose a new model with blue-green dark channel prior to remove the forward scattering and spatially dependent absorption. An example of our restoration result is shown in Figure 1. Our first contribution is that we improve the method for robustly removing global backscatter. We use RANSAC to fit the binary quadratic function from the separated background without any calibration. The second contribution is that we present a more accurate underwater propagation model with blue-green dark channel method. Figure 1. The original image captured in underwater environment (left) and the restoration result (right).