Colour reconstruction of underwater images

Objects look very different in the underwater environment compared to their appearance in sunlight. Images with correct colouring simplify the detection of underwater objects and may permit the use of visual simultaneous localisation and mapping (SLAM) algorithms developed for land-based robots underwater. Hence, image processing is required. Current algorithms focus on the colour reconstruction of scenery at diving depth where different colours can still be distinguished, but this is not possible at greater depth. This study investigates whether machine learning can be used to transform image data. First, laboratory tests are performed using a special light source imitating underwater lighting conditions, showing that the k-nearest neighbour method and support vector machines yield excellent results. Based on these results, an experimental verification is performed under severe conditions in the murky water of a diving basin. It shows that the k-nearest neighbour method gives very good results for short distances between the object and the camera, as well as for small water depths in the red channel. For longer distances, deeper water and the other colour channels, support vector machines are the best choice for the reconstruction of the colour as seen under white light from the underwater images.