Research On Image Super-resolution Algorithms Through Interpolation And Neighbor Embedding

Because of the limitation of hardware equipment,network bandwidth and storage space,the images in the devices such as cameras and on the Internet are mostly low resolution images.But the low resolution images are not clear and can not meet the need for high image resolutions,so the technology to improve the resolution of a low resolution image to obtain a high resolution image with the use of one or a few of low resolution images becomes a hot research topic in the field of image processing in recent years.The image super-resolution algorithms can be mainly classified into the following three categories:(a)the algorithms based on interpolation,(b)the algorithms based on image reconstruction,and(c)the algorithms based on learning.The algorithms based on interpolation use the values of neighboring pixels around the current pixel to be interpolated to generate the value of the current pixel,which is simple,quick and suitable for real-time applications.The algorithms based on learning build the training set to study the relationship between the low resolution blocks and the corresponding high resolution blocks to construct the high resolution blocks,which can fully use the prior knowledge of the images.Some improvements on the original algorithms based on interpolation and learning are made in the proposed algorithms.In this thesis,the proposed algorithms outperform the original algorithms.On the basis of the algorithms based on interpolation,two algorithms are proposed in this paper.They are a new interpolation-based super-resolution algorithmby the cubic spline interpolation for edge pixels and iterative update method,and a new super-resolution algorithm by interpolation in homogeneous areas.Both of the algorithms firstly extract the edge pixels and non-edge pixels by using the edge detection operator,and then the value of edge pixels and non-edge pixels are estimated with different methods.The experimental results show that the image blur caused by traditional algorithms based on interpolation can be effectively avoided in these two algorithms.In the first proposed algorithm,the edge pixels in an image are firstly detected by the Prewitt operator and then interpolated by cubic spline interpolation along the edge direction,the non-edge pixels are firstly interpolated by bi-cubic interpolation and then filtered by anisotropic operator.In the second proposed algorithm,the edge pixels are firstly detected by the Prewitt operator.Then,the homogeneous areas of edge pixels and non-edge pixels are determined.Finally,the filtering in homogeneous areas is carried out.For the algorithms based on learning,the algorithm through neighbor embedding has relatively small computation complexity and relatively good effect.On the basis of the algorithm through neighbor embedding,a three dimension super-resolution algorithm through neighbor embedding and on weighted coefficients is proposed in this paper.In this algorithm,the three dimensional discrete cosine transform(3D-DCT)coefficients are employed to extract the features of the image blocks.Then,the 3D-DCT coefficients are multiplied by a template,where the low frequency coefficients in the 3D-DCT are set with large weight values,and the high frequency coefficients are set with relatively small weight values.Simulation results verify that the reconstructed images have relatively higher quality and color consistency than conventional neighbor embedding algorithms.