Deblurring of images using blind schemes

Abstract

The thesis presents two blind deconvolution schemes for image blur removal. The two major types of blur has been worked out, namely, the gaussian blur and the motion blur. The image corrupted by the gaussian blur is reconstructed by Evolutionary algorithm using pseudo-wigner distribution. The second method deals with heuristically estimating the blur parameter of the image undergone motion blur. The gaussian effect is mostly observed in astronomical imaging. The image deblurring for motion blurred image is required due to hardware incapability of capturing the exact information of moving object or with moving camera. In this thesis, an observed image is assumed to be the two dimensional convolution of the true image with a linear-shift invariant blur, known as point-spread function, psf, and the additive noise is assumed to be zero. The Evolutionary algorithm has been implemented to remove gaussian blur. The atmospheric turbulence is mostly modelled by the gaussian psf. The algorithm proceeds by randomly generating the psf’s at each generation. The psf’s at each generation are used to estimate the true image. The best fitted images are then given as input to the next generation. After few generation, the most feasible images are chosen. These closer estimated images are fused using pseudo-wigner distribution to reconstruct the final required image. The inherent dynamic characteristic of the nature gives rise to motion blur. Whenever there is relative motion between the object to be captured and the imaging system, the image captured at that instant is suered by the type of blur known as motion blur. A new heuristic approach has been framed out with the purpose of estimating the motion blur parameter. This type of blur is characterised by its length and the motion direction. These parameters are then used to restore the image. The motion direction is estimated from the fourier domain of the observed motion blurred image. The length is then iteratively computed using Entropy and the RMSE as the quality metrics.