地球信息科学学报 ›› 2019, Vol. 21 ›› Issue (3): 315-326.doi: 10.12082/dqxxkx.2019.180229

• 地球信息科学理论与方法 • 上一篇    下一篇

基于先验HMRF的MAP分块超分重建方法

王华斌1,2,*(), 陶万成1, 李玉1, 赵泉华1   

  1. 1. 辽宁工程技术大学 遥感科学与应用研究所,阜新 123000
    2. 国家测绘地理信息局 卫星测绘应用中心,北京 100048
  • 收稿日期:2018-08-05 修回日期:2019-01-11 出版日期:2019-03-15 发布日期:2019-03-15
  • 通讯作者: 王华斌 E-mail:whb@sasmac.cn
  • 作者简介:

    作者简介:王华斌(1980- ),男,安徽合肥人,副研究员,硕士,主要从事测绘卫星数据分发、三维地理信息系统、卫星数据处理等方面的研究。E-mail: whb@sasmac.cn

  • 基金资助:
    国家自然科学基金青年基金项目(41301479);国家自然科学基金面上项目(41271435)

HMRF Prior based MAP Block Super-Resolution Reconstruction Algorithm

Huabin WANG1,2,*(), Wancheng TAO1, Yu LI1, Quanhua ZHAO1   

  1. 1. Liaoning Technical University, Institute of Remote Sensing Science and Application, Fuxin 123000, China
    2. National Bureau of Surveying and Mapping Geographic Information, Satellite Surveying Application Center, Beijing 100048, China
  • Received:2018-08-05 Revised:2019-01-11 Online:2019-03-15 Published:2019-03-15
  • Contact: Huabin WANG E-mail:whb@sasmac.cn
  • Supported by:
    National Natural Science Foundation of China Youth Fund Project, No.41301479;National Natural Science Foundation of China, No.41271435.

摘要:

针对高光谱图像应用最大后验概率(Maximum A Posteriori, MAP)超分重建后细节信息丢失严重问题,本文提出一种基于先验Huber马尔科夫随机场(Huber Markov Random Field, HMRF)模型的MAP分块超分辨率重建算法,以期提高图像超分重建质量。首先,利用主成分变换获取图像域的主要成分,在此基础上采用样条插值得到初始迭代图像;而后将初始图像域分为若干子块,在每个子块图像域上建立具有自适应阈值的HMRF模型,并结合子块图像域的保真项构建目标函数,采用梯度最快下降法求解此函数得到超分子块图像,将其重组,进而与插值后的次要成分图像相结合,最后应用主成分逆变换方法得到最终的高分辨率图像。为了验证本文算法的有效性与优越性,分别对模拟和真实图像采用本文方法和具有代表性的Tikhonov、总变分及传统HMRF模型超分重建方法进行实验对比,其中本文方法重建结果在峰值信噪比和结构相似性定量评价方面明显优于其他方法重建结果,在定性评价方面边缘结构及细节信息也更加明显,表明本文算法较为突出。

关键词: 图像域分块, 自适应阈值, 高光谱图像, HMRF模型, 主成分变换

Abstract:

The detailed information in super-resolution reconstruction of hyper-spectral image is usually lost after using the Maximum A Posteriori (MAP). To improve the quality of a reconstructed image, this paper presents a MAP block super-resolution reconstruction algorithm based on the prior Huber Markov Random Field (HMRF) model. Firstly, Principal Component Analysis (PCA) is used to obtain the main components for a given hyper-spectral image, and then the initial image is obtained by spline interpolation technique. By using main components from the PCA operation, the proposed algorithm can not only effectively reduce the usage of computation memory but also reserve most of the information from the image. After calculating the Q statistic of the initial image, it is found that stratifying the hyper-spectral image into several (e.g., seven in this study) spatial heterogeneities is an effective way to characterize the complexity of the hyper-spectral image. To this end, a suitable partitioning scheme for obtaining an optimal super-resolution reconstructed image is adopted after comparing the reconstructed results by using different blocks with different sizes. As a result, the domain of the hyper-spectral image is split into several sub-blocks. The HMRF model with an adaptive threshold is then established for each sub-block image, and an objective function is defined by combining the fidelity terms of the sub-block images. The objective function can be solved by using the gradient descent method to obtain the high resolution sub-block images, which are then combined with the interpolated secondary component images. Though some cross artifacts occur in the process, they can be removed by extending edge based methods. The effective extending edge-based method is also proposed in this paper. Finally, the final high resolution image can be obtained by using the inverse PCA operation. In order to verify the validity and the superiority of the proposed algorithm, we test the proposed algorithm, the representative Tikhonov-based algorithm, total variation-based algorithm, and the traditional HMRF model-based super-resolution reconstruction method with the simulated and real images, respectively. The testing results show that the proposed algorithm is superior to other methods in the peak signal-to-noise ratio (PSNR) and the Structure Similarity Image Measure (SSIM).The qualitative evaluation indicated that the proposed method could obtain more obvious edge structure and detailed information at the same time.

Key words: image segmentation, adaptive threshold, hyperspectral image, HMRF model, principal component transformation