基于先验HMRF的MAP分块超分重建方法
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
收稿日期: 2018-08-05
要求修回日期: 2019-01-11
网络出版日期: 2019-03-15
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
Received date: 2018-08-05
Request revised date: 2019-01-11
Online published: 2019-03-15
Supported by
National Natural Science Foundation of China Youth Fund Project, No.41301479
National Natural Science Foundation of China, No.41271435.
Copyright
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。
王华斌 , 陶万成 , 李玉 , 赵泉华 . 基于先验HMRF的MAP分块超分重建方法[J]. 地球信息科学学报, 2019 , 21(3) : 315 -326 . DOI: 10.12082/dqxxkx.2019.180229
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.
Fig. 1 Correspondence between LRH and HRH image pixel indexes图1 LRH和HRH图像像元索引对应关系 |
Fig. 2 LRH image sequence super-reconstruction process图2 LRH图像序列超分重建流程 |
Fig. 3 Simulated degradation process of hyperspectral image bands图3 高光谱图像波段的模拟降质过程 |
Tab. 1 Q statistics of different components表1 不同成分的Q统计量 |
图像 | 分层数 | Q值 | 图像 | 分层数 | Q值 | 图像 | 分层数 | Q值 |
---|---|---|---|---|---|---|---|---|
第一成分 | 4 | 0.9048 | 第二成分 | 4 | 0.9199 | 第三成分 | 4 | 0.8962 |
5 | 0.9425 | 5 | 0.9459 | 5 | 0.9312 | |||
6 | 0.9593 | 6 | 0.9670 | 6 | 0.9521 | |||
7 | 0.9701 | 7 | 0.9774 | 7 | 0.9654 | |||
8 | 0.9770 | 8 | 0.9813 | 8 | 0.9751 |
Fig. 4 Experimental results of different blocking scheme in the first major components图4 不同分块法的第一主要成分实验结果 |
Tab. 2 PSNR and SSIM for different blocking schemes of the first principal component image表2 第一主要成分图像的不同分块方案的PSNR值和SSIM值 |
未分块 | 2×2分块 | 5×5分块 | 10×10分块 | |
---|---|---|---|---|
PSNR | 19.4530 | 20.8295 | 21.9662 | 22.1437 |
SSIM | 0.8623 | 0.8779 | 0.8973 | 0.9016 |
Fig. 5 Reconstruction results of different expansion scheme in the third major components图5 第三主要成分的不同扩边方案重建结果 |
Tab. 3 PSNR and SSIM for different extension schemes表3 第三主要成分的不同扩边方案的PSNR值和SSIM值 |
无扩边 | 扩边1像元 | 扩边2像元 | 扩边3像元 | 扩边4像元 | |
---|---|---|---|---|---|
PSNR | 15.556 | 15.813 | 16.101 | 16.257 | 16.271 |
SSIM | 0.7865 | 0.8004 | 0.8115 | 0.8120 | 0.8123 |
Tab. 4 Comparison results of PSNR and SSIM表4 PSNR和SSIM数据比较结果 |
成分图像 | 质量评价 | Tikhonov | 总变分 | 传统HMRF | 本文算法 |
---|---|---|---|---|---|
第一主成分 | PSNR | 19.6534 | 20.9329 | 21.0243 | 22.1437 |
SSIM | 0.8587 | 0.8755 | 0.8766 | 0.9016 | |
第二主成分 | PSNR | 18.1071 | 18.8795 | 18.8601 | 19.7133 |
SSIM | 0.8134 | 0.8270 | 0.8312 | 0.8695 | |
第三主成分 | PSNR | 14.8539 | 15.1122 | 15.1198 | 16.2710 |
SSIM | 0.7191 | 0.7696 | 0.7653 | 0.8123 |
Fig. 6 Super-reconstruction results of the first three main components图6 前3个主要成分超分重建结果 |
Fig. 7 Simulation image super-resolution reconstruction result图7 模拟图像超分重建结果 |
Tab. 5 PSNR and SSIM of different super-resolution algorithms表5 不同超分算法的PSNR和SSIM值 |
算法 | Tikhonov | 总变分 | 传统HMRF模型 | 本文算法 |
---|---|---|---|---|
PSNR | 19.1243 | 20.2719 | 20.6865 | 22.0645 |
SSIM | 0.8565 | 0.8723 | 0.8786 | 0.8927 |
Fig. 8 Actual image super-resolution reconstruction result图8 实际图像超分重建结果 |
The authors have declared that no competing interests exist.
[1] |
[
|
[2] |
[
|
[3] |
|
[4] |
|
[5] |
[
|
[6] |
|
[7] |
|
[8] |
|
[9] |
|
[10] |
|
[11] |
|
[12] |
|
[13] |
|
[14] |
[
|
[15] |
[
|
[16] |
[
|
[17] |
[
|
[18] |
[
|
[19] |
[
|
[20] |
[
|
[21] |
|
[22] |
http://aviris.jpl.nasa.gov/.
|
[23] |
|
[24] |
[
|
[25] |
[
|
/
〈 | 〉 |