Research On SAR Image Despeckling And Optical Image Sequences Superresolution

Image quality is important for either human vision or computer vision, so that the image restoration is one of the essential topics in image processing. The widely use of the various imaging sensors brings up many new image restoration issues. In the SAR (synthetic aperture radar) images , the presence of speckle often obstacles the comprehension of the SAR images. The issue of speckle depressing has been focused for decades. For optical images, the resolution is the bottleneck problem in many applications, and the image sequences super-resolution issue becomes attractive. This thesis presents our work on these two issues.For the speckle depressing, we develop a multi-description model to describe the fundamental properties of the SAR images. Different models are used to describe different type of regions in SAR images. These models include the homogeneous region model and the non-homogeneous region model, and the later consists of the strong scatterer model, the line-like region model and the weak edge region model. Based on the model, we develop a cascading segmentation algorithm to separate the SAR image into regions, so that each of these regions can be described by one of the single models. Several sub-algorithms are developed for the cascading segmentation: the homogeneous region segmentation algorithm based on statistics hypothesis test, the line-like region segmentation algorithm and the strong scatterer segmentation algorithm. Under a unified frame, different despeckling sub-algorithms are designed to despeckle different kinds of regions: a modified ACMAP algorithm (simulated annealing correlated neighborhood MAP despeckling algorithm) is used to depress the speckle in homogeneous regions, and the macro structural information is introduced into the despeckling algorithm in heterogeneous regions to protect the structures. The proposed multi-description speckle-depressing algorithm is applied in synthetic and real SAR images, and the experimental results are prospective and satisfied. The line-like region segmentation algorithm can also be used for the line extraction in SAR images.For the image sequences super-resolution, we develop a new motion estimation algorithm in image sequences. Instead of using only two frames from the sequence, the new algorithm uses the whole sequence to estimate the motion vector of each frame, and is more accurate and robust than the former classical motion estimation algorithms. The algorithm also provides a reliability measure of the estimated result, which was introduced into the super-resolution image reconstruction to overcome the influence of the estimation error. We also establish a unified MAP frame to solve the super-resolution issue and pattern matching issue simultaneously. The frame has two advantages: first, the prior knowledge of the patterns is introduced into the super-resolution issue, and benefits the restoration of high-resolution image. Second, more information of the low-resolution sequence is used in the pattern matching. The proposed algorithms are applied in synthetic and real image sequences, and the experimental results are prospective and satisfied.