Microsoft Word 战玉丽C.doc

Size: px

Start display at page:

Download "Microsoft Word 战玉丽C.doc"

困核筹柯
4 years ago
Views:

1 3 6 Vol.3 No Journal of Computer-Aded Desgn & Computer Graphcs Jun. 09 基于图像相似性和特征组合的超分辨图像重建战玉丽,), 迟静,)*, 叶亚男,), 张彩明,,3,4) 5), 霍文远 ) ( 5004) ) ( 5004) 3) ( 500) 4) ( ) 5) ( ) (peace_world_c@6.com) 摘要 :,,., KNN ; ;,,., Wndows7 Matlab,,,, ;. 关键词 : ; ; ; 中图法分类号 : TP39.4 DOI: 0.374/SP.J Super Resoluton Image Reconstructon Based on Image Smlarty and Feature Combnaton Zhan Yul,), Ch Jng,)*, Ye Yanan,), Zhang Camng,,3,4), and Huo Wenyuan 5) ) (School of Computer Scence and Technology, Shandong Unversty of Fnance and Economcs, J nan 5004) ) (Shandong Provncal Key Laboratory of Dgtal Meda Technology, J nan 5004) 3) (Department of Software, Shandong Unversty, J nan 500) 4) (Future Intellgent Computng Collaboratve Innovaton Center, Yanta 64005) 5) (Department of Software, Nanchang Unversty, Nanchang ) Abstract: Amng at the problems of losng detaled nformaton, or generatng edge dstorton and nose whle enhancng detals n tradtonal mage reconstructon process, ths paper proposes an mage super-resoluton reconstructon method based on mage cross-scale smlarty and feature combnaton. Usng mage cross-scale smlarty, we frst apply KNN algorthm to respectvely establsh the mappng relatonshp of pxel features and gradent features between the hgh and the low resoluton mages. Then, we use the mappng relaton of pxel features to reconstruct the hgh-resoluton mage contanng hgh-frequency nformaton from the nput mage. Fnally, we use sngular value thresholdng to obtan the effectve : ; : 基金项目 : (677309); (ZR08JL0); (J8RA7); (U6098);. 战玉丽 (99 ),,, ; 迟静 (980 ),,,,,, ; 叶亚男 (994 ),,, ; 张彩明 (955 ),,,,, CAGD&CG ; 霍文远 (997 ),,.

2 6, : 09 hgh-frequency nformaton of the nput mage, and use the gradent feature mappng relatonshp to enlarge the hgh-frequency nformaton whch wll be supermposed onto the hgh-resoluton mage to generate the fnal mage reconstructon result. We utlze the mage segmentaton database of UCLA as the expermental data and dsplay the expermental results n the Matlab software under Wndows 7. The expermental results show that our method can reconstruct mages wth rch texture detals, clear edges, and sgnfcantly enhanced mage detals. It greatly mproves vsual effects and obectve ndcators. Moreover, our method does not need to rely on external databases. Key words: super resoluton; detal enhancement; cross-scale smlarty; sngular value decomposton,,., [], () () (). (). (Bcubc).., [-3],,., []. ().,, [4]., Wang [5],. [6],,,. (3).,.,,., Yang [7],,,. Dong [8],.. Zhao [9],. Ta [0],,. Tmofte [],,., Tmofte [],.,, Dong [3],.,,. Freeman [4], ;,,.,,,. Ebrahm [5],. [6-7], Chang [6],, (locally lnear embeddng, LLE) K,. [8-0].,, ;,. Freedman []

3 00 3,,, ;,,,.,,.,, ;., ;,, ;,.. 本文方法. ().. (). (),,,. (3).,, KNN.,,. (4). (3) HI HB, HI HB., 3 3,.,,.,... 获得图像高频信息.,,. : I, M ;

4 6, : 0 M, I y ; I y, 3 3,, ;, B... 高频信息的奇异值阈值化,,,. (sngular value thresholdng, SVT), B, B '. () B( m n), m n B, 0 B U V 0 0, U m m; V n n ; dag(,,, m ), (,,, m) B, 3 m 0. (),, ( ). B,, 0.,,,, m, D dag{( ) }, D B. (3). () D B '. a. m n, D n n, ( ) ( ) ( ) 0,, 0 n, ; B T B' UD ( ) V,, U' U m n. ' b. m n, D m m, T. ( ) ( ) ( k ) 0,, 0 k m, k ; B T B' UD V',, T V' V T m n. c. m n, D m n, ( ) ( ) ( d ) 0,, 0 d m n, d ; B T B ' =UD V..3 高低分辨率图像之间的特征关系训练 [0].,,., I, HI, ; B', HB, HB HI.,,,.. I, Y ; I M, M y. Y y.. (feature constraned nterpolaton, FCI) [],,.,.,,, [7, 6], [3]. [4],,. Zhao [5],., KNN Y y, Y

5 0 3 y. : I, KNN Y y R, I, HI ; Y, y, Y, y Y y., 3 3,., Y { Y }, y { y }.,,,, N ;,,, n. Y Y, y y. Y Y, y { y }, KNN S { y } n (), 0 N, n y Y { y } K., () mn Y S () S Y, { y }, n y y (3) arg mn Y Yy y y S. y S, 0 ;, e Yy y S Yy y., y Y. I, 3 3,, P. y, I y., y, I I, I y y,,,, n, n I. : B ', KNN Y y R, R B ', HB, h [,0,,0,] v [,0,,0,]. Y T y, Y y,, 33, Y y. Y { Y }, y y { }.,,,, N ;,,, n ; Y Y, y y. Y y ;, Y () y Y, S { y } ny.,,,, N;,,, n ; n y Y { y } K. Y, y, S Y, S (3), S. B', 3 3,,, Q. B' y, B' y,, B' B' y y,,,, n, n B'..4 重建结果图像 () Y y R, HI. S,,,, N G, G HI,. 3b baby HI. () Y y R, HB. S,,,, N G, G HB,. 3c baby HB. (3) HI HB, 3 3,, HH. 3 baby ( 0).,,..5 本文方法步骤 :. I.. HH. Step., I M, y,

6 6, : 03 3 baby I Y. Step. Y, y R. Y, y R. Step3. B, SVT, B', 3 3,, Q. Step4. B' I, P, R R P Q HI HB. Step5. HI HB, HH. 实验结果和分析, Bcubc Yang [7] Freedman [] Dong [8] A+SelfExSR [6] SRCNN [3]., SRCNN (ImageNet)*,.. Wndows 7, AMD Radeon RX550, Intel(R) Core(TM) GHz, Matlab 06a. 3. (peak sgnal to nose rato, PSNR), (structural smlarty ndex, SSIM),., PSNR SSIM., 3 Y, C b, C r Y, PSNR. Y, Y PSNR SSIM. Y, RGB YC b C r, Y., C b, C r Bcubc, Y., Y.,, Dong.. 主观数据对比 4~ 0.,,,. 4~ 6, Bcubc, ; Yang [7], 5 6 ; Freedman [], ; Dong [8], 4 ; A+ [], SelfExSR [6], SRCNN [3], 4. 5, A+ [] SRCNN [3],.,,

4 a. b. Bcubc c. Yang [7] d. Freedman [] e. 5

7 04 3 a. b. Bcubc c. Yang [7] d. Freedman [] e. Dong [8] f. A+ [] g. SelfExSR [6] h. SRCNN [3]. 4 a. b. Bcubc c. Yang [7] d. Freedman [] e. Dong [8] f. A+ [] g. SelfExSR [6] h. SRCNN [3]. 5

8 6, : 05 a. b. Bcubc c. Yang [7] d. Freedman [] e. Dong [8] f. A+ [] g. SelfExSR [6] h. SRCNN [3]. 6 a. b. Bcubc c. Yang [7] d. Freedman [] e. Dong [8] f. A+ [] g. SelfExSR [6] h. SRCNN [3]. 7

06 3 4,, 5. 7, ; Dong [8],,. ; A+ [] SRCNN [3],, ; SelfExSR [6],.

Yang [7] d. Freedman [] e. Dong [8] f. A+ [] g. SelfExSR [6] h. SRCNN [3].

9 06 3 4,, 5. 7, ; Dong [8],,. ; A+ [] SRCNN [3],, ; SelfExSR [6],. 8~ 9, Bcubc ; Yang [7] ; Freedman [],, ; Dong [8] 9, ; A+ [] SelfExSR [6] SRCNN [3] 9. 0, Bcubc Freedman [], Freedman [] a. b. Bcubc c. Yang [7] d. Freedman [] e. Dong [8] f. A+ [] g. SelfExSR [6] h. SRCNN [3]. 9 a. b. Bcubc c. Yang [7] d. Freedman [] e. Dong [8] f. A+ [] g. SelfExSR [6] h. SRCNN [3]. 8 ; Yang [7],, ; Dong [8] ; A+ [] SelfExSR [6] ; SelfExSR [3],.. 客观数据对比,, ~ 3.,,. 3, 3a, 3c 3d.,, ~3,,., 0~70.

739 3.48 8.4 4.534 4.6 4.403 8.57 35.490 3.860 9.846 36.55 3.440 3.64 3.440 35.964 6.678 4.594 3.849 6.589 4.5 4.438 4.407 7.45 3.538 8.48 6.886 3.78 7.

10 6, : 07 a. b. Bcubc c. Yang [7] SSIM d. Freedman [] e. Dong [8] f. A+ [] a. =5 b. =70 g. SelfExSR [6] h. SRCNN [3]. 0 c. =40 d. =70 3 PSNR PSNR SSIM 8,., PSNR SSIM Dong [8] SSIM Bcubc, ; Dong [8] Bcubc,.,,. 表 8 种方法 PSNR 比较 db Bcubc Yang [7] Freedman [] Dong [8] A+ [] SelfEx- SR [6] SRCNN [3]

11 08 3 表 8 种方法 SSIM 比较 Bcubc Yang [7] Freedman [] Dong [8] A+ [] SelfEx- SR [6] SRCNN [3] 结语. ;,, ;,,,, ;,,.,,,. 参考文献 (References): [] Pu Jan, Zhang Junpng, Huang Hua. A revew of super resoluton algorthms[j]. Journal of Shandong Unversty: Engneerng Scence, 009, 39(): 7-3(n Chnese) (,,. [J]. :, 009, 39(): 7-3) [] L X, Orchard M T. New edge-drected nterpolaton[j]. IEEE Transactons on Image Processng, 00, 0(0): 5-57 [3] Zhou F, Yang W M, Lao Q M. Interpolaton-based mage super-resoluton usng multsurface fttng[j]. IEEE Transactons on Image Processng, 0, (7): [4] Baker S, Kanade T. Lmts on super-resoluton and how to break them[j]. IEEE Transactons on Pattern Analyss and Machne Intellgence, 00, 4(9): [5] Wang L F, Xang S M, Meng G F, et al. Edge-drected sngle-mage super-resoluton va adaptve gradent magntude self-nterpolaton[j]. IEEE Transactons on Crcuts and Systems for Vdeo Technology, 03, 3(8): [6] Marquna A, Osher S J. Image super-resoluton by TV-regularzaton and Bregman teraton[j]. Journal of Scentfc Computng, 008, 37(3): [7] Yang J C, Wrght J, Huang T S, et al. Image super-resoluton va sparse representaton[j]. IEEE Transactons on Image Processng, 00, 9(): [8] Dong W S, Zhang L, Lukac R, et al. Sparse representaton based mage nterpolaton wth nonlocal autoregressve modelng[j]. IEEE Transactons on Image Processng, 03, (4): [9] Zhao Y, Wang R G, Ja W, et al. Local patch encodng-based method for sngle mage super-resoluton[j]. Informaton Scences, 08, : [0] Ta Y W, Lu S C, Brown M S, et al. Super resoluton usng edge pror and sngle mage detal synthess[c] //Proceedngs of the IEEE Conference on Computer Vson and Pattern Recognton. Los Alamtos: IEEE Computer Socety Press, 00: [] Tmofte R, De Smet V, Van Gool L. A+: Adusted anchored neghborhood regresson for fast super-resoluton[j] Anchored neghborhood regresson for fast example-based super-resoluton[c] //Proceedngs of the IEEE Internatonal Conference on Computer Vson. Los Alamtos: IEEE Computer Socety Press, 03: [] Tmofte R, de Smet V, van Gool L. A+: adusted anchored neghborhood regresson for fast super-resoluton[c] //Proceedngs of the th Asan Conference on Computer Vson. Hedelberg: Sprnger, 05: -6 [3] Dong C, Loy C C, He K M, et al. Image super-resoluton usng deep convolutonal networks[j]. IEEE Transacton on Pattern Analyss and Machne Intellgence, 06, 38(): [4] Freeman W T, Jones T R, Pasztor E C. Example-based super-resoluton[j]. IEEE Computer Graphcs and Applcatons, 00, (): [5] Ebrahm M, Vrscay E R. Solvng the nverse problem of mage zoomng usng self-examples [C] //Proceedngs of the 4th Internatonal Conference Image Analyss and Recognton. Hedelberg: Sprnger, 007: 7-30 [6] Chang H, Yeung D Y, Xong Y M. Super-resoluton through neghbor embeddng[c] //Proceedngs of the IEEE Computer

12 6, : 09 Socety Conference on Computer Vson and Pattern Recognton. Los Alamtos: IEEE Computer Socety Press, 004: 75-8 [7] Zhao Y, Wang R G, Wang W M, et al. Multlevel modfed fnte radon transform network for mage upsamplng[j]. IEEE Transactons on Crcuts and Systems for Vdeo Technology, 06, 6(): [8] Shechtman E, Iran M. Matchng local self-smlartes across mages and vdeos[c] //Proceedngs of the IEEE Conference on Computer Vson and Pattern Recognton. Los Alamtos: IEEE Computer Socety Press, 007: -8 [9] Glasner D, Bagon S, Iran M. Super-resoluton from a sngle mage[c] //Proceedngs of the IEEE Internatonal Conference on Computer Vson. Los Alamtos: IEEE Computer Socety Press, 009: [0] Yang J C, Ln Z, Cohen S. Fast mage super-resoluton based on n-place example regresson[c] //Proceedngs of the IEEE Conference on Computer Vson and Pattern Recognton. Los Alamtos: IEEE Computer Socety Press, 03: [] Freedman G, Fattal R. Image and vdeo upscalng from local self-examples[j]. ACM Transactons on Graphcs, 0, 30(): Artcle No. [] Zhang Xn, Zhang Fan, L Xueme, et al. Feature constraned mult-example based Image Super-resoluton[J]. Journal of Computer-Aded Desgn & Computer Graphcs, 06, 8(4): (n Chnese) (,,,. [J]., 06, 8(4): ) [3] Chan T M, Zhang J P, Pu J, et al. Neghbor embeddng based super-resoluton algorthm through edge detecton and feature selecton[j]. Pattern Recognton Letters, 009, 30(5): [4] Yang Xn, Fe Shumn, Zhou Dake, et al. Image fast super-resoluton reconstructon based on class predctor and degradaton model[j]. Journal of Southeast Unversty: Natural Scence Edton, 03, 43(): 35-38(n Chnese) (,,,. [J]. :, 03, 43(): 35-38) [5] Zhao Z, He X F, Ca D, et al. Graph regularzed feature selecton wth data reconstructon[j]. IEEE Transactons on Knowledge and Data Engneerng, 06, 8(3): [6] Huang J B, Sngh A, Ahua N. Sngle mage super-resoluton from transformed self-exemplars[c] //Proceedngs of the IEEE Conference on Computer Vson and Pattern Recognton. Los Alamtos: IEEE Computer Socety Press, 05:

Microsoft Word - 系统建设1.doc

Microsoft Word - 系统建设1.doc 计算机系统应用 http://www.c-s-a.org.cn 20 年第 20 卷第 4 期一种高性能近红外光人脸检测和眼睛定位算法张昌明, 童卫青, 王燕群 ( 华东师范大学计算机科学技术系, 上海 20024) 摘要 : 提出了一种新型的近红外光人脸检测和眼睛