【作者】 张明；

【导师】 胡访宇；

【作者基本信息】 中国科学技术大学 ， 通信与信息系统， 2010， 硕士

【摘要】 图像超分辨率重建(Super-resolution Reconstruction, SRR)是指从降质低分辨率图像序列中构造出高分辨率图像的分辨率增强技术。本文对图像超分辨率重建算法以及关键插值算法问题进行了研究。随着对高分辨率的需求,低成本获取装置产生的图像需要通过信号和图像处理技术来提高质量,因此超分辨率重建已成为图像研究领域热点之一。其中,高精度的运动配准算法、盲重建算法、稳定的高效算法一直是超分辨率重建研究领域的重难点。首先,本文介绍了超分辨率重建的数学模型,对几种经典超分辨率重建算法做了讨论。重建算法主要分为两类:频域法和空域法。现今广泛应用的图像超分辨率重建算法如:贝叶斯法、凸集投影法、混合MAP-POCS法、迭代反投影法等。其次,低分辨率图像必须插值到高分辨率图像网格,所以本文研究了图像插值理论。多项式内插:最近邻插值、线性插值、双三次插值具有固有的局限性,因此引进了样条插值和核回归插值。最后,仿真实验的图像结果展示了各算法的重建质量。本文引进正则化项解决超分辨率重建反问题的病态性。利用现今广泛采用2阶范数作为正则化的基础上,在经典核回归方法重建图像中,提出的新型权重函数与取样窗范围内随机像素点位置有关。然后引入局部方向信息,利用可控核回归法,进一步提高重建图像的质量。仿真结果表明,改进后核回归的方法不仅提高了图像的质量,而且降低了重建图像的均方误差。更多还原

【Abstract】 Image super-resolution reconstruction (SRR) refers to a resolution enhancement technology, which produce the image of high-resolution (HR) image from a set of degraded low-resolution (LR) images. This dissertation investigates image SRR algorithm and the key issues of interpolation algorithm.With the demand of HR, the images produced from low cost acquisition system need the signal and image processing technology to improve the quality, so SRR has become one of the hot areas of the research on image. Among this, the high-precision movement registration algorithm, the blind reconstruction algorithms, and stable of fast and the effective algorithm are the focus of the difficulties of SRR.In this paper, Firstly, the mathematical model of SRR is introduced, and several classical SRR algorithms are discussed. Frequency algorithms and spatial algorithms are the two main reconstruction technique. Now, Bayesian approach, projections onto convex sets (POCS) Approach, MAP-POCS hybrid approach and iterative back projection (IBP) approach are wide used in SRR algorithms. Secondly, LR image must be inserted into the HR image grid, so image interpolation theory is investigated. Polynomial interpolation, as nearest neighbor interpolation, bi1inear interpolation, cubic interpolation, has the inherent drawbacks, so we adopt spline Interpolation, kernel regression interpolation. Finally, simulation results of the HR image show the quality of there algorithm.We adopt the regularization term to solve the ill-posed inverse problem of SRR. With the current wide spread use of norm 2 as the regularization, classical kernel regression (CKR) method was utilized for image reconstruction. We propose the novel weight function related to the position of the random pixels in the scale of the sample window. Then steering kernel regression (SKR) method contained local orientation was used to improve the quality of the reconstruction image. Simulation results demonstrated the improvement of image quality and the reduction of root mean square error (RMSE).更多还原