一种基于MAP的超分辨率图像重建的快速算法  

A Novel Fast Algorithm for MAP Super-Resolution Image Reconstruction

作　　者：肖创柏[1] 禹晶[1] 薛毅[2] 

机构地区：[1]北京工业大学计算机学院,北京100124 [2]北京工业大学应用数理学院,北京100124

出　　处：《计算机研究与发展》2009年第5期872-880,共9页Journal of Computer Research and Development

基　　金：北京市自然科学基金项目(4092006)~~

摘　　要：超分辨率图像重建技术就是通过融合多幅变形、模糊、有噪、频谱混叠的低分辨率降质图像(或视频序列)来重建一幅高质量高分辨率图像.MAP估计算法是一种广泛使用的统计重建方法.针对标准MAP估计算法运算量大的问题提出了两点改进.第1点是当计算梯度时直接计算目标函数的增量,避免了函数值的冗余计算;第2点是采用非精确一维搜索确定步长,避免了运算量庞大的海塞矩阵的计算.实验结果表明,提出的改进在保持重建效果基本不变的前提下,在很大程度上提高了MAP超分辨率图像重建方法的速率,与此同时保证了算法的收敛性.Super-resolution image reconstruction has recently drawn considerable attention within the research area. For some special-purpose imaging devices such as medical imaging, remote sensor imaging and video capturing, the acquired images cannot often achieve a higher resolution because of the limitation of imaging mechanism and imaging sensor. Super-resolution image reconstruction methods attempt to create a single high-resolution and high-quality image from multiple low-resolution observations （or a video sequence） degraded by warping, blurring, noise and aliasing. So far, existing super-resolution methods are all confronted with the problem of slow convergence and expensive computation. To satisfy the requirement of real-occasion applications, a fast super-resolution reconstruction algorithm is built upon the MAP framework. In the proposed algorithm, two improvements are presented to reduce the high computational complexity of the standard MAP algorithm. The first improvement is to compute directly the increment of the MAP objective function as the component of the gradient vector, which avoids the redundant computation of the objective function. The second one is to select the Armijo rule to identify the step size, which avoids the computation of the computationally demanding Hessian matrix. Experimental results show that the computation time is reduced significantly, whereas the solution＇s convergence is guaranteed and the similar quality is maintained.

关 键 词：图像重建 超分辨率 MAP(最大后验概率) 梯度投影算法 一维搜索 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]