Research On The Methods Of Blind Image Restoration

Image restoration method is used in many applications, such as space navigation, remote sensing, astronomy, medical imaging and forensic science. Most of restoration approaches require the prior knowledge of blur and some features of noise. However, such an assumption may not be satisfied in practice. Recovering an original image from several degraded versions without such knowledge is called as blind image restoration. Since blind image restoration doesn't depend on PSF (point spread function), it has extensive use. This dissertation intensively explores blind image restoration method.This dissertation specifies the basic methods of blind image restoration, analyzes the degradation model of image and the ill-posed problem during restoration.Based on original nonnegativity and support constrains recursive inverse filtering (NAS-RIF) algorithm, an improved blind image restoration algorithm is proposed in this dissertation. In order to control the trade-off between fidelity to the observed image and smoothness of the restored image and to prevent noise amplification, a newly cost function of the NAS-RIF algorithm can be obtained by adding space-adaptive terms and a regularization term. Analyzing the method of choosing regularization operator, we get the conclusion: the regularization parameter should be high-pass. Based on the theory of constrained least square, different parameters are used to restore degraded images, the results show that: the regularization parameter should be chosen up on the properties of degraded image and noise variance. An estimating method of noise variance is proposed in this paper. The regularization parameter and space-adaptive terms can be calculated through the estimated noise variance and the local variance of the degraded image. This improved algorithm uses conjugate-gradient routine to calculate the optimal result. The improved algorithm and original algorithm are both used to restore three different degraded images. The experimental results show that: the results of△SNR are increased by 0.2073db, 1.0239db, 2.8628db compared with original NAS-RIF algorithm. It demonstrates that the improved algorithm is more efficient.Due to wavelet transform exhibits very good localization properties in frequency and in spatial domain. An improved NAS-RIF algorithm for blind image restoration based on wavelet transform is presented in this dissertation. The degraded image is decomposed to obtain its wavelet coefficients in wavelet domain. NAS-RIF algorithm is used to restore degraded image in each sub-bands, different regularization terms are used in different sub-bands. Estimating the noise variance in each sub-band, the adaptive regularization parameters can be calculated through the local properties of the observed image and the noise variance. The experimental results show that this blind image restoration method is more efficient than traditional space-adaptive regularization method.