联合深度置信网络与邻域回归的超分辨率算法

doi:10.13232/j.cnki.jnju.2019.04.008

南京大学学报(自然科学版) ›› 2019, Vol. 55 ›› Issue (4): 581–591.doi: 10.13232/j.cnki.jnju.2019.04.008

所属专题：测试专题

联合深度置信网络与邻域回归的超分辨率算法

黄炜钦^1,²(),高凤强^1,²,陈俊仁^1,²,李婵¹

1. 厦门大学嘉庚学院信息科学与技术学院，漳州，363105
2. 嘉庚学院微柏工业机器人创新实验室，漳州，363105

收稿日期:2019-05-28 出版日期:2019-07-30 发布日期:2019-07-23
通讯作者: 黄炜钦 E-mail:wqhuang@xujc.com
基金资助:
厦门大学嘉庚学院校级科研孵化项目(2018L03);福建省高校杰出青年科研人才培育计划(闽教科[2018]47号)

Image super⁃resolution algorithm joint deep belief network and neighborhood regression

Weiqin Huang^1,²(),Fengqiang Gao^1,²,Junren Chen^1,²,Chan Li¹

1. School of Information Science & Technology, Xiamen University Tan Kah Kee College, Zhangzhou, 363105, China
2. Tan Kah Kee College?VIBOT Industrial Robot Innovation Lab, Zhangzhou, 363105, China

Received:2019-05-28 Online:2019-07-30 Published:2019-07-23
Contact: Weiqin Huang E-mail:wqhuang@xujc.com

摘要/Abstract

摘要：

为了提高重建的质量和速度，提出一种联合深度置信网络与邻域回归的超分辨率算法.一方面，结合字典学习与神经网络表示的联系对传统的深度置信网络进行调整，采用该网络模型实现字典学习，充分利用该模型突出的学习能力，使字典具有更好的特征表达能力，从而提高图像的重建质量.另一方面，在基于字典学习的超分辨率框架中融入邻域回归思想.首先，利用最近邻域算法确定字典原子的最近邻域映射关系；然后以此为基础，结合邻域回归方法，离线计算高、低分辨率投影矩阵；最后在重建过程中将该投影矩阵应用于图像重建.该方法避免了字典学习中的系数求解过程，降低了计算的复杂度，提高了重建的速度.实验表明，算法具有更高的峰值信噪比和结构相似度，同时极大地提高了图像的重建速度.

关键词: 超分辨率, 字典学习, 深度置信网络, 邻域回归

Abstract:

To improve the quality and speed of reconstruction,an image super?resolution algorithm joint deep belief network and neighborhood regression is proposed. For one thing,combine the relationship between dictionary learning and the neural network representation to adjust the traditional deep belief network and apply it to learn the dictionary,that make full use of the outstanding learning ability of the deep belief network to make the dictionary have the better ability to express feature. As a result,the quality of image reconstruction can be improved. For another,we integrate the idea of neighborhood regression in the framework of super?resolution based on dictionary learning. First of all,we determine the nearest neighborhood mapping relationship of the dictionary atom with the nearest neighborhood algorithm,and then based on it,the high resolution and low resolution projection matrix can be calculated offline using the neighborhood regression. Finally,the projection matrix can be applied to reconstruct images during the reconstruction process. This method avoids the process of solving coefficients in dictionary learning and reduces the computational complexity. Consequently, It has the ability to improve the speed of reconstruction.The experiments show that the algorithm has a higher Peak Signal?to?Noise Ratio and Structural Similarity,and it also increases the speed of image reconstruction greatly.

Key words: super?resolution, dictionary learning, deep belief network, neighborhood regression

中图分类号:

TP391

黄炜钦, 高凤强, 陈俊仁, 李婵. 联合深度置信网络与邻域回归的超分辨率算法[J]. 南京大学学报(自然科学版), 2019, 55(4): 581–591.

Weiqin Huang, Fengqiang Gao, Junren Chen, Chan Li. Image super⁃resolution algorithm joint deep belief network and neighborhood regression[J]. Journal of Nanjing University(Natural Sciences), 2019, 55(4): 581–591.

图/表 16

图1

图2

图3

图4

图5

图6

图7

表1

图8

图9

表2

图10

图11

图12

图13

表3

参考文献 20

1	潘宗序，禹晶，肖创柏等. 基于自适应多字典学习的单幅图像超分辨率算法. 电子学报，2015，43(2)：209-216.
	Pan Z X，Yu J，Xiao C B，et al.Single image super resolution based on adaptive multi?dictionary learning. Acta Electronica Sinica，2015，43(2)：209-216.
2	苏衡，周杰，张志浩. 超分辨率图像重建方法综述. 自动化学报，2013，39(8)：1202-1213.
	Su H，Zhou J，Zhang Z H.Survey of super?resolution image reconstruction methodsActa Automatica Sinica，2013，39(8)：1202-1213.)
3	FreemanW T，JonesT R，PasztorE C. Example?based super?resolution. Computer Graphics and Applications IEEE，2002，22(2)：56-65.
4	YangJ C，WrightJ，HuangT，et al. Image super?resolution as sparse representation of raw image patches∥Proceedings of 2008 IEEE Conference on Computer Vision and Pattern Recognition. Anchorage，AK,USA：IEEE Press，2008：1-8.
5	ZeydeR，EladM，ProtterM. On single image scale?up using sparse?representations∥Proceeding of 7^th International Conference on Curves and Surfaces. Springer Berlin Heidelberg，2010：24-30.
6	WoldS，EsbensenK，GeladiP. Principal component analysis. Chemometrics and Intelligent Laboratory Systems，1987，2(1-3)：37-52.
7	TimofteR，DeV，Van GoolL. Anchored neighborhood regression for fast example?based super?resolution∥Proceedings of 2013 IEEE International Conference on Computer Vision. Sydney,Australia：IEEE，2013：1921-1927.
8	DongC，LoyC C，HeK M，et al. Image super?resolution using deep convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence，2014，38(2)：295-307.
9	DonohoD L. For most large underdetermined systems of equations，the minimal 1?norm near?solution approximates the sparsest near?solution. Communications on Pure and Applied Mathematics，2006，59(7)：907-934.
10	EnganK，AaseS O，HusoyJ H. Frame based signal compression using method of optimal directions (MOD)∥Proceedings of the 1999 IEEE International Symposium on Circuits and System. Orlando,FL,USA：IEEE，1999：1-4.
11	AharonM，EladM，BrucksteinA. K?SVD：an algorithm for designing overcomplete dictionaries for sparse representation. IEEE Transactions on Signal Processing，2006，54(11)：4311-4322.
12	KeyvanradM A，HomayounpourM M. A brief survey on deep belief networks and introducing a new object oriented toolbox (DeeBNet). 2014,arXiv：1408.3264.
13	SmolenskyP. Information processing in dynamical systems：foundations of harmony theory. Boston：MIT Press，1986,194-281.
14	KrizhevskyA. Learning multiple layers of features from tiny images. Master Dissertation. Toronto：University of Toronto，2009.
15	HintonG E，OsinderoS，TehY W. A fast learning algorithm for deep belief nets. Neural Computation，2014，18(7)：1527-1554.
16	蒋建国，陈亚运，齐美彬等. 基于自相似性和稀疏表示的图像超分辨率重建. 光电工程，2015，42(12)：74-81.
	Jiang J G，Chen Y Y，Qi M B，et al.Image super?resolution reconstruction based on selfsimilarity and sparse representation. Opto?Electronic Engineering，2015，42(12)：74-81.
17	BevilacquaM，RoumyA，GuillemotC，et al. Low?complexity single image super?resolution based on nonnegative neighbor embedding∥Proceedings of British Machine Vision Conference. Guildford,British：BMVA Press，2012：1-10.
18	李祚林，李晓辉，马灵玲等. 面向无参考图像的清晰度评价方法研究. 遥感技术与应用，2011，26(2)：239-246.
	Li Z L，Li X H，Ma L L，et al.Research of definition assessment based on no?reference digital image quality. Remote Sensing Technology & Application，2011，26(2)：239-246.
19	路鹤晴，朱国英，卓维海等. 医用X射线CT辐射剂量影响因素研究. 上海医学影像，2008，17(2)：93-96，136.
	Lu H Q，Zhu G Y，Zhuo W H，et al.An evaluation on influence factors of radiation dose for the medical X?ray computed tomography. Shanghai Medical Imaging，2008，17(2)：93-96，136.
20	ClarkK，VendtB，SmithK，et al. The cancer imaging archive (TCIA)：maintaining and operating a public information repository. Journal of Digital Imaging，2013，26(6)：1045-1057.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed

测试图像	PSNR/SSIM	Bicubic	L1SR	SISR	ANR	SRCNN	DBNNR
Baby	PSNR	33.91	34.29	35.08	35.13	35.01	35.26
Baby	SSIM	0.9244	0.9226	0.9402	0.9415	0.9401	0.9425
Brid	PSNR	32.58	34.11	34.57	34.60	34.91	34.95
Brid	SSIM	0.9422	0.9530	0.9615	0.9623	0.9623	0.9643
Head	PSNR	32.88	33.17	33.56	33.63	33.55	33.70
Head	SSIM	0.8386	0.8382	0.8572	0.8601	0.8581	0.8614
Barbara	PSNR	26.25	26.39	26.76	26.69	26.66	26.80
Barbara	SSIM	0.7813	0.7890	0.8081	0.8078	0.8076	0.8128
Face	PSNR	32.82	33.12	33.53	33.62	33.58	33.68
Face	SSIM	0.8374	0.8374	0.8567	0.8599	0.8583	0.8612
Foreman	PSNR	31.18	32.05	33.19	33.23	33.35	33.75
Foreman	SSIM	0.9231	0.9247	0.9426	0.9431	0.9442	0.9465
Ppt3	PSNR	23.71	24.98	25.23	25.03	25.32	25.56
Ppt3	SSIM	0.8874	0.9025	0.9204	0.9123	0.9274	0.9293
Zebra	PSNR	26.63	27.95	28.49	28.43	28.87	29.07
Zebra	SSIM	0.8361	0.8648	0.8775	0.8794	0.8826	0.8874
平均值	PSNR	30.00	30.76	31.30	31.30	31.40	31.60
平均值	SSIM	0.8713	0.8790	0.8955	0.8958	0.8976	0.9007

测试图像库	L1SR	SISR	SRCNN	DBNNR
Set5	23.51	1.02	3.44	0.30
Set14	52.72	2.11	10.00	0.62

无参考指标	平均梯度	信息熵	Brenner函数	能量函数
原始图	2.9443	6.6701	4261738	3379712
DBNNR	3.6527	6.7088	6087568	5284806

[1]	罗鸣威2，王怀登1*，丁　尧1，袁　杰2. 变焦序列图像超分辨率重建算法研究[J]. 南京大学学报(自然科学版), 2017, 53(1): 165-.
[2]	王喆正, 唐晔, 杨育彬 * . 利用图像类标信息的自调式字典学习方法 [J]. 南京大学学报(自然科学版), 2015, 51(2): 320-327.

联合深度置信网络与邻域回归的超分辨率算法

Image super⁃resolution algorithm joint deep belief network and neighborhood regression

RichHTML

PDF (PC)

摘要/Abstract

引用本文

使用本文

图/表 16

参考文献 20

相关文章 2

Metrics

本文评价

推荐阅读 8