Underwater image restoration: Super-resolution and deblurring via sparse representation and denoising by means of marine snow removal
Underwater imaging has been widely used as a tool in many fields, such as marine industry, deep-sea mining, aquaculture and water assessment. However, a major issue is the quality of the resulting images and videos. Due to the light's interaction with water and its constituents, the acquired underwater images and videos often suffer from a significant amount of scatter (blur and haze) and noise. Furthermore, since data transmission from the equipment mounted under water to the station above water is still a challenge, usually a compressed and low-resolution version of the data is transferred. In the light of these issues, this thesis considers the problems of low-resolution, blurred and noisy underwater images and proposes several approaches to improve the quality of such images/video frames. This is undertaken through two main contributions. The first major contribution of this work is the super-resolution and deblurring of single underwater images. This is done by using a set of compact high and low-resolution cluster dictionaries where sparse representation is used as the regularizer. Since such an approach inevitably calls for a model selection criterion in both learning and reconstruction stages, a scaleinvariance model is proposed to properly establish the link between the low and high-resolution feature spaces. The subject of the second major contribution is image denoising. Besides additive noises such as sensor noise, the visibility in underwater images is reduced by the presence of suspended particles in water. This represents an unwanted signal, which is also disruptive for advanced computer vision tasks, such as segmentation. Since this phenomenon is a real signal and part of the scene, two-fold approaches consisting of first detection and then removal of such particles, are proposed. To avoid the uncertainty introduced by using local information for restoration, some global priors of the scene are learned, which are then used to estimate the parts of the scene that are covered by the particles. For this, a Gaussian-based background subtraction approach is proposed to obtain static features of the scene. These are used as training data for learning the priors. Quantitative and qualitative experiments conducted over real and simulated underwater images and video frames validate the success of the proposed approaches at improving the image resolution and deblurring image features significantly as well as detecting and removing marine particles, while the object edges are preserved.
Rostock, Univ., Diss., 2018