Research On Image Interpolation Based On The Local Filtering Model

Image interpolation is an outstanding digital image processing technology to obtain a high resolution(HR) image from its low resolution(LR) counterpart, which is widely used in security surveillance, military reconnaissance, medical imaging, and so on.The thesis introduces image interpolation and some classical interpolation algorithms briefly. Then, the local filtering model based image interpolation methods are mainly focused. The local filtering model is characterized by expressing one pixel using its local neighbors, including the auto-regressive(AR) model and the filter model. The main work of this thesis is as follows.Firstly, an improved image interpolation method by AR model and soft-decision estimation is proposed, including training phase and interpolation phase. In training phase, a dictionary is built according to local patch feature and its optimal regularization weight. In interpolation phase, the weight of LR pixels in a local window is obtained through their similarity to the central pixel, and the piece-wise auto-regressive(PAR) model parameters are then estimated by the weighted least squares. Moreover, the optimal regularization weight is gotten through searching the nearest sample in dictionary. The missing pixels are finally reconstructed by soft-decision estimation. Furthermore, the PAR parameters and the interpolated pixels are both corrected using Expectation-Maximization algorithm. The proposed method performs better in reserving edge information.Secondly, a novel image interpolation method through estimating the bilateral filtering parameter is proposed, whose motivation is that the more accurate filtering parameter can be estimated, including rough estimation phase and precise estimation phase. In rough estimation phase, the principal components analysis is used to estimate the edge direction of local region, then the gotten direction is quantized into four parts, including 0°, 45°, 90° and 135°, according to human visual feature. Moreover, different directional parameters are used to estimate the current range distance. In precise estimation phase, the structure similarity of all HR range distance within the semi- local region is measured on the initial HR image. Then, the current range distance is re-calculated through maximum likelihood estimation by the corresponding similar observations. Experimental results show that the proposed method can effectively improve the interpolated image.