Method Studies In Super Resolution Image Restoration

Super Resolution image restoration (SR) is a newcome image processing method that extracting higher resolution images containing more details from an image sequence of lower resolution, by image process such as motion estimation, deblurring and denoising. Currently SR is one active image processing field, in which the obtained algorithms can be used to improve image resolution in remote sensing and medical applications. In this paper, algorithms in three kinds of SR processing manner are studied, that are: only considering SR interpolation, considering together blurring estimation and high-resolution image restoration, and simultaneously considering image motion, blurring estimation and high-resolution image restoration.When only considering SR interpolation, some improvement to the existing wavelet-based SR interpolation is proposed. It is pointed out that because of the adopted 2-D real Discrete Wavelet Transformation(DWT) lacking of shift-invariance and direction selectivity, thus in the restored high resolution image obtained by Wavelet-based SR interpolation, there will appears some undesired Ringing Effects, and only few directions of details can be selected to restore by wavelet-based SR interpolation. Dual-Tree Complex Wavelet (DT CWT), which has nearly shift-invariance and direction selectivity with limited data redundancy, is proposed to replace 2-D real DWT in wavelet-based SR interpolation, thus not only make more selectivity of detail directions, but can also reduce the Ringing Effects in the restored high resolution image. Related computation steps are given. Tests results on synthetical lower resolution image sequences are also presented to show the validation of the proposed SR interpolation based on DT CWT.When considering image blurring estimation and high resolution image restoration together, improvements are proposed to the regularization method of the existing linear regularized and parametric SR blind restoration (LR PBSR). It is analyzed that because in the existing LR PBSR the adapted Tikhonov regularization is linear shift-invariance, and the adapted identity regularization operation is high-pass and low-not-stop, thus over smoothing will be induced into edge areas of the restored image which will decimate the restored image. In the same linear Tikhonov regularization frame, By adding into the regularization term a high-pass and low-stop Laplacian regularization operator and a weight matrix defined by the gradient vector magnitude which representing the local smoothing features, an local adopted linear regularization and parametric blind SR is realized, which make the regularization...