Image Restoration using Machine Learning
Inhaltsverzeichnis
Reference
Asmatullah Chaudhry: Image Restoration using Machine Learning. Thesis, Faculty of CS and Engineering, Topi, Pakistan, March 2007.
DOI
Abstract
Restoration of degraded images has become an important and effective tool for many technological applications like space imaging, medical imaging and many other post-processing techniques. Most of the image restoration techniques model the degradation phenomena, usually blur and noise, and then obtain an approximation of the image. Whereas, in realistic situation, one has to estimate both the true image and the blur from the degraded image characteristics in the absence of any a priori information about the blurring system. The objective of this thesis is to develop a new punctual kriging based image restoration approach using machine-learning techniques. To achieve this objective, the research concentrates on the restoration of images corrupted with Gaussian noise by making good tradeoffs between two contradicting properties; smoothness versus edge preservation. This thesis makes the following contributions:
(1) Quantitative analysis of the at hand punctual kriging based image restoration technique is carried out,
(2) Fuzzy logic, punctual kriging and fuzzy averaging are used intelligently to develop a better image restoration technique,
(3) A new image quality measure is proposed in terms of the semi-variograms to judge the performance of image restoration techniques,
(4) Analysis of the effect of neighbourhood size on negative weights and the subsequent improvement in punctual kriging based image restoration is performed,
(5) To avoid both the problems of matrix inversion failure and the negative weights in punctual kriging, artificial neural network is used to develop a neuro-fuzzy filter for image denoising,
(6) Further, using genetic programming, a hybrid technique for image restoration based on fuzzy punctual kriging is developed, the developed machine learning technique uses local statistical measures along with kriged information for subsequent pixel estimation.
Main parameters considered for evaluation of the proposed technique are image quality measure and computational cost. The image quality measures used for evaluation and comparison include MSE, PSNR, SSIM, wPSNR, VMSE and VPSNR. A series of empirical investigations have been made to evaluate the performance of the proposed techniques using database of standard images that show the effectiveness of our methodology.
Extended Abstract
Bibtex
Used References
Links
Full Text
http://prr.hec.gov.pk/thesis/2056.pdf (no c&p)
