Institute of Information Theory and Automation

Image Deconvolution Methods using Prior Information

Project leader: Prof. Ing. Jan Flusser, DrSc.
Department: ZOI
Supported by (ID): ASCR and CSIC (Spain) bilateral
Duration: 2009 - 2010
Publications at UTIA: list

Abstract:

The project proposal belongs to the area of digital image processing and deals with sophisticated methods for image enhancement/restoration. It is a continuation of our current joint project “Mathematical Methods for Resolution Enhancement of Digital Images” and two previous projects “Multifocus and multimodal image fusion techniques for biomedical applications” and “New developments on multimodality image fusion and spatial variant image processing”, which all have been very successful. They led to several joint publications (see the list above), fruitful short-term visits, and attracted several PhD. students that joined our research groups. Current reconstruction methods, which remove blur (Point Spread Function – convolution kernel) from images, provide satisfying results if multiple images of the same scene (objects) are available. If the images contain slightly different blurs then multichannel blind deconvolution methods, can estimate the blurs and the original sharp image in a fairly stable manner. We have substantially contributed to this field, which several our publications demonstrate. Most of them were results of collaboration with the Spanish group in the frame of previous three bilateral projects. Often it is costly or even impossible to acquire more than one image of the same scene. Then we face a single-channel blind deconvolution problem, which lacks satisfying solution in the general case. The nature of this problem renders any general algorithm to be extremely unstable (ill-posed) and the only way how to achieve a stable solution is to incorporate prior knowledge of blurs and images. Such priors are commonly described by probability density functions or regularization terms, which play the role of constraints that force the solution to lie in a set of admissible blurs and images. Recently several attempts to define better priors appeared in the literature, which suggest that this tantalizing problem is becoming more and more popular. It is unrealistic to assume that some general priors can be built, since this is in contradiction with the role of priors. They must be tailored to the particular family of images and blurs we are dealing with, in order to be of any use. Our new proposal is concerned with novel forms of prior information, which can be used for image deconvolution. We plan to elaborate on different definitions of priors based on an application in hand. In order to clarify our goal, we give two examples from different application fields. In medical imaging, PET images (volumes) provide important functional information, however the image quality is poor due to severe noise and blur. Modern tomography scanners can also generate CT images simultaneously. CT is of much higher resolution and quality, but provides only anatomical (not functional) information. We propose using the information from CT to build priors for the PET image and achieve better reconstruction of the functional image. We envisage that this can be a great asset to nuclear medicine. In photography, photos taken from hand are sometimes blurred due to motion of the camera. To reconstruct the original image one has to have good priors for blurs. Taking another photo with much short exposure time results in a sharp but noisy and underexposed image. We can use information from both images to build blur priors and then perform reconstruction of the blurred single image.
Responsible for information: ZOI
Last modification: 18.04.2014
Institute of Information Theory and Automation