基于MAP的POCS算法实现图像超分辨率复原

发布时间：2018-04-21 22:22

本文选题：凸集投影 + 超分辨率复原　；参考：《吉林大学》2017年硕士论文

【摘要】：图像超分辨率复原技术是利用一幅低分辨率图像或者多幅低分辨率图像之间的信息互补,运用相应的复原算法来获得一幅比原始图像细节更多、视觉效果更好的高分辨率图像。目前国内外的超分辨率复原技术,一般使用的都是序列图像复原,即去除图像序列的降质因素之后,再将多幅图像融合为一幅高分辨率、高质量图像。图像的超分辨率复原技术发展到现在,已经广泛应用于计算机视觉、图像压缩、医学图像处理、遥感图像与军事图像捕获采集、数字电视转换、生物特征识别等诸多领域。纵观国内外的研究现状,图像超分辨率复原方面的研究已经取得了一定的成果,尤其是使用较多的凸集投影法和最大后验概率算法。但是目前已有的算法还是存在结果不唯一、边缘保持能力不足、图像细节丢失等问题。针对目前存在的这些问题,为了进一步提高超分辨率复原图像的质量,本文对凸集投影超分辨率复原算法和最大后验概率算法进行改进,创新点以及研究内容主要包含以下三个方面:①提出了基于修正点扩散函数的凸集投影超分辨率复原算法。为了保持图像的边缘特性,提高抗噪能力,对传统的点扩散函数加一个权值因子进行修正,将点扩散函数细分到八个方向,使边缘方向的点扩散函数不变,边缘正交方向的权值因子减小。从而达到了增强图像边缘保持能力,提高复原图像质量的效果。②提出了改进的最大后验概率超分辨率复原算法。这一部分的重点内容在于概率模型的选取。不同于传统的方式,本文中先验概率模型选择吉布斯随机场模型,并且为了尽量多的保留图像的细节信息,我们对模型中吉布斯项的系数进行了修正。条件概率模型选择Gauss-Markov随机场模型,为了缓解过度平滑导致图像边缘信息丢失的问题,本文中首先提取出图像边缘信息,对图像进行边缘增强处理。然后将图像加入改进的最大后验概率算法中进行超分辨率复原,从而进一步增加图像的细节信息。③提出了基于最大后验概率的凸集投影超分辨率复原算法。将两种算法进行结合,充分抑制二者缺点,发挥各自优势。首先利用最大后验概率算法,对几幅低分辨率待处理图像进行估计,得出一幅初始的高分辨率图像。将这幅图像作为凸集投影超分辨率复原的初始参考图像,并对其利用八方向Sobel算子和多方向形态滤波器进行预处理,之后加入第一部分提出的基于修正点扩散函数的凸集投影超分辨率复原算法中进行迭代运算。通过能量有界性等先验信息求得凸集相对应的投影算子。利用改进的凸集投影法约束迭代求解,直到达到设定的阈值或达到设定的迭代次数,退出迭代,得出最终的超分辨率复原图像。利用软件仿真,得出不同图像在不同算法下的超分辨率复原结果,利用主观评价方法以及客观评价中的峰值信噪比、均方误差和结构相似度等多个指标,从多方面对复原图像质量进行评估。实验结果表明,本文提出的算法可以有效提高超分辨率复原图像的质量。
[Abstract]:The image super resolution restoration technology uses a low resolution image or a number of low resolution images to complement each other, and uses the corresponding restoration algorithm to obtain a high resolution image with more details and better visual effect than the original image. Like restoration, after removing the degradation factor of the image sequence, multiple images are fused into a high resolution, high quality image. The super resolution restoration technology of images has developed to the present. It has been widely used in computer vision, image compression, medical image processing, remote sensing images and military image capture and acquisition, digital television conversion, biology. There are some achievements in the research of image superresolution restoration, especially the convex set projection method and the maximum posterior probability algorithm. However, the existing algorithms still have not only the results, the edge retention ability is insufficient, the image details are lost, and so on. In order to further improve the quality of super-resolution restoration images, this paper improves the convex set projection superresolution restoration algorithm and the maximum posterior probability algorithm in order to improve the quality of super-resolution restoration image. The innovation points and the research content mainly include the following three aspects: (1) the projection superdivision of convex set based on the modified point diffusion function is proposed. In order to keep the edge characteristics of the image and improve the anti noise ability, the traditional point diffusion function is modified by adding a weight factor, the point diffusion function is subdivided into eight directions, the point diffusion function of the edge direction is changed, the weight value of the edge orthogonal direction is reduced, thus the edge retention ability of the image is enhanced. To improve the quality of restoration image. Secondly, an improved maximum posterior probability superresolution algorithm is proposed. The key content of this part is the selection of the probability model. It is different from the traditional way. In this paper, the prior probability model of this paper selects the Gibbs random field model, and in order to keep the details of the image as much as possible, we are right The coefficient of the Gibbs term in the model is modified. The conditional probability model selects the Gauss-Markov random field model. In order to alleviate the problem of the image edge information loss caused by excessive smoothness, the image edge information is extracted and the edge enhancement processing is carried out in this paper. Then the image is added to the improved maximum posterior probability algorithm. Super-resolution restoration to further increase the details of the image. Thirdly, a convex set projection super-resolution restoration algorithm based on the maximum posterior probability is proposed. The two algorithms are combined to fully suppress the shortcomings of the two ones and give full play to their respective advantages. An initial high resolution image is obtained. This image is used as the initial reference image of the projection superresolution restoration, and its eight direction Sobel operator and multi direction morphological filter are preprocessed. Then the first part is added to the convex set projection super-resolution restoration algorithm based on the modified point diffusion function. Iterative operation. The projection operator corresponding to the convex set is obtained through the priori information of the energy boundedness. The improved convex set projection method is used to constrain the iterative solution, until the set threshold is reached or the set iteration number is reached, and the final super-resolution restoration image is drawn out. By using the subjective evaluation method, the peak signal to noise ratio in the objective evaluation, the mean square error and the structural similarity, the quality of the restored image is evaluated from many aspects. The experimental results show that the proposed algorithm can effectively improve the quality of the super-resolution image.

【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】