Research On Image Super-resolution Reconstruction Based On Non-local Similarity

Spatial resolution refers to the pixel density in an image. The higher the resolution is, the more image details are. There are two ways to obtain high-resolution images. One is improving the optical element and sensor. The other is using the signal processing method to improve the resolution. However, improving the optical element and sensor cost too much which is not fit for the common business application, and the size of pixel in sensor can't be reduced infinitely. So researchers pay more attention on the signal processing method.In this paper, we research on the field of super-resolution reconstruction. We propose a novel super resolution (SR) construction algorithm and its fast method. The experimental results show our proposed method can improve the output result both in visual perception and in objective estimation. We have devoted to the below several aspects:First, we study the iterative back-projection algorithm (IBP). This algorithm is a classical SR method and has low computational complexity, which can be applied in real time applications. We analyze the weakness of the IBP method and propose to take advantage of the image non-local similarity to reconstruct high resolution image. The proposed method can remove the artifacts of the reconstruction image of IBP.Second, we study the bilateral filtering and non-local mean algorithms. To strengthen the true image edges, we filter the initial estimate image using bilateral filter in the pre-processing step. In the post-processing procedure we learn the structural content of low resolution pixel and correlation among the pixels with similar structure. Then we use the correlation to guide the output image reconstructed by the IBP algorithm.Third, the proposed algorithm has high computational complexity. To satisfy the real time applications, we present the fast algorithm, which has been shown at below:a) Classifying the pixels. We classify the pixels in accordance with the property of the coordinates. In the local area, the pixels with the same coordinate property can share the same motion vectors.b) Edge detection based on adaptive threshold. The proposed method is mainly to improve the edge of the image. So we process the edges not all the image in the post-processing step. We propose a simple edge detection method based on the adaptive threshold.c) Reducing the dimensionality using principle component analysis (PCA). During the block match process, the operations of high dimensional data cost too much time, and the data has too much redundance. We transfer the high dimensional data to the low dimensional data using the PCA technique, which can improve the speed of the algorithm.