Spectral Based Blur Classification and Parameter Estimation Approaches for Image Restoration

Abstract

Extracting information from a blurred image without any prior knowledge of image, its blurring and sensing mechanism is of great interest in image processing and vision based systems. The image restoration process requires blur type classification and extracting parameters of sensing mechanism. One approach is spectral based blur classification in blind image deconvolution. Identifying the blur kernel, also known as point-spread-function (PSF) and then restoring image using non-blind methods has many solutions in various domains – spatial, spectral and transform, but they are complex and give limited quality results. Blind deconvolution requires joint estimate of PSF and original image, which becomes iterative or recursive, and incorporate some prior knowledge about process. This thesis presents Blur identification mechanisms using novel spectral approach that is invariant based, Random Forest classifier based using skewness and kurtosis parameters, and novel application of MNIST CNN architecture. This thesis also proposes spectral based new framework for estimation of motion and defocus blur parameters. Motion blur parameter estimation is based on innovative formulation of trigonometric relation between spectral lines to relate blur length and angle. The MNIST CNN architecture is applied for motion blur parameter estimation. In this thesis, four methods are proposed for defocus blur parameter estimation. First technique estimates the defocus blur radius from the radius of spectral nulls of the blurred image using the proposed First White method. Second method estimates the defocus radius by determining the signature of radius of the spectral nulls of the blurred image. The third method formulates a polynomial relation between radius of this spectral null and actual radius of defocus blur. The fourth method employs random forest classifier to estimate the blur radius using a proposed set of features. The combination of random forest classifier and proposed feature set adds novelty to defocus parameter estimation approach. All the proposed approaches for blur type classification and parameter estimation for blurred images were experimented on large number of images from standard datasets like USCSIPI dataset, Berkeley segmentation dataset and Pascal VOC dataset. The best blur classification accuracy achieved from all the three proposed techniques is around 97% and the classification results validate the ability of proposed approaches in distinguishing blur type. The best attained blur parameter estimation accuracy of proposed methods is nearly 100% for motion blur and around 98% for defocus blur. All the proposed approaches are also compared with other earlier reported methods. The parameters, thus estimated by the all the proposed approaches prove effective in efficient restoration of known and unknown blurred images. Along with better classification and parameter estimation accuracies, the key contributions of the thesis are simplification of radius estimation algorithm, novel use of statistical parameters (skewness and kurtosis) in Random Forest for blur classification and defocus radius estimation, and novel application of MNIST CNN architecture for blur classification and motion parameter estimation.