Studies On Image Super Resolution Reconstruction

In many electronic imaging applications such as infrared imaging system and CCD, the image resolutions of the video sequences are limited to the array densities of the sensors. Moreover, the pixel difference of the optics, atmosphere and system noise will blur and warp the images. Though we can enhance the resolutions through changing the high precise optics or increasing the chip size, it will make a high cost and suffer some practical limitations in the fact. Therefore, it is important and economical to enhance the resolution by super-resolution (SR) reconstruction technique.SR reconstruction is to produce high quality, high-resolution (HR) images from a set of degraded, nonidentical, low-resolution (LR) images. With the development of the recent years, there presented many useful algorithms such as the frequency algorithms, non-uniform interpolation algorithms and spatial iterative algorithms, which are various for characters and benifets. Moreove, SR reconstruction of compressed signals differtiates from that of uncompressed signals and suffers some new difficulties and challenges for introducing the quantization noise.In the dissertation, we study the relative techniques of the SR reconstruction and analyze the theoretical basis and technologies extensively. In order to enhance the reconstructed properties for SR reconstruction, the dissertation presents some new algorithms to focuses on the stability, computational expense and the reconstruction of the techniques.First, we analyze the stability of the SR reconstructed algorithm. The chapter two presents a new model for SR reconstruction based on the statistical SR theory. The MAP estimation is guaranteed to be well posed by choosing the prior distribution as a strict convex function. The edge image is modeled as a Lorentzian distribution to reconstruct, which is iterated by the Lorentzwidth parameter adaptively. Experiments demonstrate the validity and superiority of the Lorentzian model for SR reconstruction. The proposed Lorentzian-based MAP algorithm can get ideal and stable reconstructions for the resolutions.Then, the dissertation studies MAP SR constructed problems for compressed images and videos. Since the quantization of DCT coefficients introduces errors in the image representation and a loss of signal information, the chapter three introduces two ways to reduce the artifacts of the MAP optimizations in SR reconstruction for the compressed videos, respectively.From the point of view of the conditional probability density function (PDF), we incorporate the prior knowledge of DCT coefficients into modeling the quantization noise. Instead of modeling the uniform distribution for the quantization noise, we calculate the DCT covariance matrix of the noise based on the Laplacian distribution of AC DCT coefficients. As a result, the spatial convariance matrix is exploited to construct a general MAP SR framework.On the other hand, we fuse the SR reconstruction and the post-processing problem to model the prior image as piecewise Markov random field (MRF). After classifying pixels by the variances, we classify the pixels as smooth, texture, edges and ringing properties. Different indicator functions are constructed for the different cliques, which aims to enhance the resolution and remove the artifacts of the reconstructed image while preserving high-frequency information.At last, the chapter four proposes a new wavelet-based interpolation for the multiple images to enhance the resolution. The algorithm fuses the multiframe information by using the cycle-spinning (CS) methodology and the relative motion information between the LR images. The algorithm incorporates the non-uniform interpolation and restoration process of the reconstructed algorithms, consequently, it solves the limitations of the traditional non-uniform interpolations and can be applicable to reconstruct the HR image when the degraded models are identical across all LT images with low computational expense.