Parameters Estimation For Image Restoration

Abstract

Image degradation generally occurs due to transmission channel error, camera mis-focus, atmospheric turbulence, relative object-camera motion, etc. Such degradations are unavoidable while a scene is captured through a camera. As degraded images are having less scientific values, restoration of such images is extremely essential in many practical applications. In this thesis, attempts have been made to recover images from their degraded observations. Various degradations including, out-of-focus blur, motion blur, atmospheric turbulence blur along with Gaussian noise are considered. Basically image restoration schemes are based on classical, regularisation parameter estimation and PSF estimation. In this thesis, five different contributions have been made based on various aspects of restoration. Four of them deal with spatial invariant degradation and in one of the approach we attempt for removal of spatial variant degradation. Two different schemes are proposed to estimate the motion blur parameters. Two dimensional Gabor filter has been used to calculate the direction of the blur. Radial basis function neural network (RBFNN) has been utilised to find the length of the blur. Subsequently, Wiener filter has been used to restore the images. Noise robustness of the proposed scheme is tested with different noise strengths. The blur parameter estimation problem is modelled as a pattern classification problem and is solved using support vector machine (SVM). The length parameter of motion blur and sigma (σ) parameter of Gaussian blur are identified through multi-class SVM. Support vector regression (SVR) has been utilised to obtain a true mapping of the images from the observed noisy blurred image. The parameters in SVR play a key role in SVR performance and these are optimised through particle swarm optimisation (PSO) technique. The optimised SVR model is used to restore the noisy blurred images. Blur in the presence of noise makes the restoration problem ill-conditioned. The regularisation parameter required for restoration of noisy blurred image is discussed and for the purpose, a global optimisation scheme namely PSO is utilisedto minimise the cost function of generalised cross validation (GCV) measure, which is dependent on regularisation parameter. This eliminates the problem of falling into a local minima. The scheme adapts to degradations due to motion and out-of-focus blur, associated with noise of varying strengths. In another contribution, an attempt has been made to restore images degraded due to rotational motion. Such situation is considered as spatial variant blur and handled by considering this as a combination of a number of spatial invariant blurs. The proposed scheme divides the blurred image into a number of images using elliptical path modelling. Each image is deblurred separately using Wiener filter and finally integrated to construct the whole image. Each model is studied separately, and experiments are conducted to evaluate their performances. The visual as well as the peak signal to noise ratio (PSNR in dB) of restored images are compared with competent recent schemes.