Restoration Algorithms for Blurred and Noisy Images

Abstract

Images are generally degraded due to faulty sensor, channel error, camera mis-focus, atmospheric turbulence, relative object-camera motion etc. Such conditions are inevitable while capturing a scene through a camera. As degraded images are of little scientific values, restoration of such images is utmost essential. In this thesis, investigations have been made to recover images from their degraded observations.Degradations like impulsive noise, out-of-focus blur, and motion blur are considered in isolation. Also we deal with blur and Gaussian noise together in one of our contribution.Two algorithms are proposed to detect random valued impulsive noise in a contaminated image. These algorithms are based on back propagation network (BPN) and support vector machine (SVM). The identified noisy pixels are then filtered using a novel filtering mechanism, which is based on the spatial property, the radiometric property, and the impulsive strength of the test pixel.The parameter of out-of-focus blur is modeled as a pattern recognition problem and is solved using multi-class SVM, whose performance is in turn enhanced by the use of binary decision tree.Support vector regression (SVR) is used to map a blurred image to its original version and two schemes are proposed in this regard. In one of the schemes, local variance is utilized to cluster the input data set to reduce the training time. In the second scheme, SVR is modeled through fuzzy systems and the image restoration problem is solved with greater efficiency.The classic method of iterative blind deconvolution is improved by incorporating more constraints and a stopping criterion. This has resulted in termination of the algorithm in fewer terations and also with improved results, which are visually more appealing.To deal with blur and Gaussian noise together, an adaptive regularized scheme is proposed.Here a regularization parameter is replaced with a regularization function.In addition, an anisotropic regularization operator is used to deal with smoothing operation of the objective function.Each model is studied separately and experiments are conducted to evaluate their performances. The restoration measures are compared with competent schemes.Keywords: Image restoration, random valued impulsive noise, out-of-focus blur, motion blur,SVM, SVR, multi-class SVM, blind image deconvolution, regularization.