Seeing Through Water
Several attempts have been lately proposed to tackle the problem of recovering the original image of an underwater scene using a sequence distorted by water waves. The main drawback of the state of the art is that it heavily depends on modelling the waves, which in fact is ill-posed since the actual behavior of the waves along with the imaging process are complicated, and include several noise components; therefore, their results are not satisfactory. In this paper, we revisit the problem by proposing a data-driven two-stage approach, each stage is targeted toward a certain type of noise. The first stage leverages the temporal mean of the sequence to overcome the structured turbulence of the waves through an iterative robust registration algorithm. The result of the first stage is a high quality mean and a better structured sequence; however, the sequence still contains unstructured sparse noise. Thus, we employ a second stage at which we extract the sparse errors from the sequence through rank minimization. Our method converges faster, and drastically outperforms state of the art on all testing sequences even only after the first stage.
Figure 1 shows the main steps in our method for seeing through water. First, we iteratively register the frames to their mean while updating the mean at every iteration. At every registration iteration, we estimate the blur level of the mean and adjust the blur of the frames to match the blur level of mean. When the registration is finished, the remaining distortion is spare; therefore, we extract the sparse errors through rank minimization.
Results – Sample Frames
Figure 2 shows the reconstruction results for sample frames from each sequence after applying each stage of our algorithm. The first stage overcomes most water turbulence; however, sparse errors are only eliminated after the second stage. The Final two rows show the reconstructed images and the sparse errors respectively after rank minimization.
Evolution of the Mean
Figure 3 shows the evolution of the mean in stage 1. Left to right: The mean after each iteration of registration. Top to bottom: Stage 1 applied without blurring or deblurring, with mean deblurring, and with our frame blurring. After three iterations, the mean is significantly enhanced in all cases. However, underwater words on the left part of the image like “Imaging”, “Water”, “Scene”, “Tracking”, and “Reconstruction” are only correctly reconstructed using the frame blurring.
Figure 4 shows the image restoration results on standard sequences from . The first column shows a sample frame from the input video, which is severely distorted. The second column shows the temporal mean of the sequence. The third column is the result from . Finally our results are shown in the last two columns, after the first and second stages respectively. Results from our method clearly outperform  on all sequences even after the first stage.
Omar Oreifej, Guang Shu, Teresa Pace, and Mubarak Shah. A Two-Stage Reconstruction Approach for Seeing Through Water,International Conference on Computer Vision and Pattern Recognition (CVPR) 2011.