Research On Reconstruction Algorithm Of Super Resolution For Image Sequences

In the acquisition process of digital image, there often exists many factors to influencethe imaging systems, and leads to resolution of the image not to meet the applicationrequirements. However, due to hardware fabrication complexity, how to increase the currentresolution level by applying software tools has become a topic of great interest. In thetraditional restoration problem only a single input image is available for processing, whichresult in the limited capability of resolution improving. Then a new technique, superresolution reconstruction is required.This paper focuses on the super resolution reconstruction for image sequences. Thepaper mainly studies the approach to obtain a high resolution image from observed multiplewarped, blurred, decimated and noisy low resolution images. The main research works areas follows:First, sub-pixel image registration algorithms were achieved. Image registration wasthe first step for super resolution reconstruction. A high resolution image could be obtainedonly when the registration accuracy was sub-pixel. An optical flow method on a Gaussianpyramid structure was utilized in this paper with iteration Lucas-Kanada algorithm to get theresidual optical flow. This algorithm was compared with Keren algorithm and Vandewallealgorithm. Experimental results showed that the optical flow method could gain wellsub-pixel accuracy.Second, the algorithm of MAP super resolution reconstruction was improved. Therelative contributions of each low resolution images were assumed equal in the traditionalalgorithm, which made the result of reconstruction be influenced. Based on the traditionalalgorithm, the proposed method introduced new channel weights to adjust the relativecontributions adaptively. Experimental results showed that even if the structure of the lowresolution images were different, the algorithm could still provide better reconstructingresults.Third, a robust super resolution reconstruction algorithm was proposed. The regularterm in the proposed algorithm was based on orthogonal gradient operator. Displacementoperator was introduced to achieve the regular term and the regularization parameter wasdetermined adaptively. In order to deal with model error, the proposed algorithm wasimplemented on the framework of L1-norm. Experiments results showed that the proposed algorithm could effectively restrained the Gaussian noise and Salt and Pepper noise. At thesame time, the reconstruction time was reduced sharply.Fourth, a spatial adaptive super resolution reconstruction algorithm was proposed. Thenew algorithm gave full consideration to local characteristic of the image, and introducedspatial adaptive weighting matrix. By combining global regularization parameter with localregularization matrix, the new algorithm overcame regularization error and noiseamplification error which were generated by regularization. The results of the experimentsindicated that the proposed algorithm effectively reduced the error, and protected the detailof the image.