Diploma thesis: Multi-View Deconvolution of Biomedical Images
Thesis by: Moritz Blume
Advisor: Prof. Nassir Navab
Supervision by: Darko Zikic
Biomedical images are often degraded during image acquisition. Two very common degradations that occur in many biomedical imaging modalities are blur and noise. Corrupted images are more difficult to interpret by the physician or researcher and so the goal is to restore the original image.
Blurring can be mathematically modeled by convolution, and within this model the blur is characterized by the point-spread-function (PSF). Two kinds of blur can be distinguished: shift-invariant and shift-variant blur. Shift-invariant blur is constant over the whole image, and shift-variant blur is characterized by a PSF that is dependent on the spatial position in the image.
The process of restoring a blurred image is referred to as deconvolution in the literature.
There are two major classes: non-blind deconvolution algorithms that assume that the PSF is known in advance and blind deconvolution algorithms that do not need any a priori knowledge regarding the PSF.
We propose a new blind deconvolution method. As an input, our method takes several observations of the same scene, each of them acquired from a different point of view, and their respective spatial transformations. The transformations are either directly known by dedicated image acquisition systems, or they can be retrieved by image registration or tracking. The output is the restored original image and PSF.
This multi-view setting occurs frequently in practice. However, most of the methods available in the literature do not specifically take advantage of the information provided by this setting.
We perform a quantitative analysis of our new method and compare it other deconvolution methods. There are two main contributions: Firstly, our algorithm is able to perform unparameterized blind deconvolution in the shift-variant case. To the best of our knowledge, known methods for blind shift-variant deconvolutions are dependent on an a priori blur model. Secondly, in the shift-invariant case, our algorithm allows for a better restoration at higher noise levels than comparable methods.