الفهرس 
INTRODUCTIONOnly 14 pages are availabe for public view
 |  | from | 130 |  |  |
| | | from | 130 | | |
Abstract
Image restoration is the procedure of recovering a clear image from an initially blurred image. The blurred image indicates a degraded version of the main scene. This degradation may occur because of atmospheric distortions, illumination and color imperfections because of camera exposure, noise and blur. Blur is often the dominant degradation cause in image processing applications.
Blur is a deterministic image degradation that causes a significant reduction in the visual quality of digital images. It affects the identification and extraction of the beneficial information in the recorded image.
With the rapid advance in computer vision systems and applications, it becomes imperative to detect, analyze and restore blurred regions in degraded images.
In this thesis, we propose a new spatially-varying motion and defocus blur detection and restoration method. Blur is detected from a single image without requiring any prior knowledge about the parametric blur kernel, the camera settings, or information about the input image.
The proposed blur detection approach detects both motion and defocus blurred regions in partially blurred images. Firstly, blur-type classification is performed by using Hough transform and the structured learning edge-detection method.
Secondly, the initial blur kernel is estimated with a kernel direction that maximizes the likelihood of a local window being blurred by incorporating either vertical or positive diagonal kernels. Blur detection is implemented using a kernel specific feature.
Thirdly, initial blur image regions are refined with the support of the reduced edges image segmentation (CCP) method. Finally, neighboring information is utilized to refine detected blur results.
For image restoration both motion and defocus blur kernel are estimated, and the initial blur regions are detected using a kernel-specific feature. Subsequently, initial blur regions are refined with the support of the contour-guided color palette (CCP) image segmentation method and neighboring information. Finally, blurred regions are recovered using approximated kernels.
The performance of the proposed approach is verified using various images from public data set. The experimental results validate the efficiency of the proposed approach in detecting blur in both motion and defocus blurred images.
The proposed method attains anoverall blur detection accuracy improvement (across datasets I and II) between 4.8% and 18.4% overall comparative methods.The proposed method in general is with the best performance and achieves over 77.6% precision for low recall compared to 71.2% and 74.1 using Liu et al. and Chakrabarti et al., respectively.