Please wait a minute...
Front. Inform. Technol. Electron. Eng.  2016, Vol. 17 Issue (1): 32-40    DOI: 10.1631/FITEE.1500171
Original article     
Image meshing via hierarchical optimization
Hao XIE,Ruo-feng TONG()
Institute of Artificial Intelligence, Zhejiang University, Hangzhou 310027, China
Download: HTML     PDF(1027KB)
Export: BibTeX | EndNote (RIS)      

Abstract  

Vector graphic , as a kind of geometric representation of raster images, has many advantages, e.g., definition independence and editing facility. A popular way to convert raster images into vector graphics is {image meshing}, the aim of which is to find a mesh to represent an image as faithfully as possible. For traditional meshing algorithms, the crux of the problem resides mainly in the high non-linearity and non-smoothness of the objective, which makes it difficult to find a desirable optimal solution. To ameliorate this situation, we present a hierarchical optimization algorithm solving the problem from coarser levels to finer ones, providing initialization for each level with its coarser ascent. To further simplify the problem, the original non-convex problem is converted to a linear least squares one, and thus becomes convex, which makes the problem much easier to solve. A dictionary learning framework is used to combine geometry and topology elegantly. Then an alternating scheme is employed to solve both parts. Experiments show that our algorithm runs fast and achieves better results than existing ones for most images.



Key wordsImage meshing      Hierarchical optimization      Convexification     
Received: 27 May 2015      Published: 05 January 2016
CLC:  TP391.7  
Fund:  the National Natural Science Foundation of China(No. 61170141);National High-Tech R&D Program (863) of China(No. 2013AA013903)
Cite this article:

Hao XIE,Ruo-feng TONG. Image meshing via hierarchical optimization. Front. Inform. Technol. Electron. Eng., 2016, 17(1): 32-40.

URL:

http://www.zjujournals.com/xueshu/fitee/10.1631/FITEE.1500171     OR     http://www.zjujournals.com/xueshu/fitee/Y2016/V17/I1/32


Image meshing via hierarchical optimization

Vector graphic , as a kind of geometric representation of raster images, has many advantages, e.g., definition independence and editing facility. A popular way to convert raster images into vector graphics is {image meshing}, the aim of which is to find a mesh to represent an image as faithfully as possible. For traditional meshing algorithms, the crux of the problem resides mainly in the high non-linearity and non-smoothness of the objective, which makes it difficult to find a desirable optimal solution. To ameliorate this situation, we present a hierarchical optimization algorithm solving the problem from coarser levels to finer ones, providing initialization for each level with its coarser ascent. To further simplify the problem, the original non-convex problem is converted to a linear least squares one, and thus becomes convex, which makes the problem much easier to solve. A dictionary learning framework is used to combine geometry and topology elegantly. Then an alternating scheme is employed to solve both parts. Experiments show that our algorithm runs fast and achieves better results than existing ones for most images.

Fig. 1 Inputs and outputs. Given an image, 'Lena' (a), and the sampling ratio (1% in this example), our algorithm generates a mesh-based vector graphic (d) with its rendered result (b). Detail regions of both (a) and (b) are enlarged and shown in (c), in which upper and lower parts correspond to (a) and (b), respectively. References to color refer to the online version of this figure
1: HierarchyEstablishment
2: MeshInitialization
3: $i\Leftarrow N - 1$
4: while Level i ≥ 0 do
5: ????ObjectiveConvexification(i)
6: ????AlternatingOptimization(i)
7:????$i\Leftarrow i - 1$
8: end while
Algorithm 1 Pipeline of our algorithm
Fig. 2 Hierarchy establishment: a hierarchy with N levels, from coarser (higher level) to finer (lower level), is established. At each level, an image of this level and a mesh from the last level are employed to generate a mesh of this level. Note that the mesh of level 0 is the desired final mesh
Fig. 3 Image meshing in a 1D image: the original problem (a) and that after convexification (b). References to color refer to the online version of this figure
Image SR (%) PSNR (dB)
Adams(2011) Xieet al.(2014) Ours
Lena127.375928.860029.7244
229.826330.954431.3691
330.980131.991832.1422
Fruits125.926926.958927.7363
228.532729.442429.7518
329.962930.802430.5960
Flowe125.802826.971827.7477
227.921729.607129.9863
328.661530.800930.9495
Falcon127.319629.966630.4886
228.532331.972932.1160
328.867532.787732.7106
Car123.164524.647825.0299
225.694126.885727.2032
326.816327.949028.0324
Fish123.987325.230125.9045
226.629627.763628.1316
328.107429.216629.2648
Leaf126.806127.881628.6272
228.839529.984630.3554
329.592431.000831.1287
Swan124.482625.961426.7547
226.989028.378328.9821
328.365129.701930.0601
Frog127.440428.977629.8903
229.884831.058431.5725
330.879131.968932.2853
Table 1 Peak signal-to-noise ratio (PSNR) comparison among our algorithm and the algorithm proposed by Adams (2011) and Xie et al. (2014)
Fig. 4 Wireframes: (a) is the mesh generated from Adams (2011), employed as the initial mesh; (b) is ours. As shown, the triangles near edges are skinnier after optimization of our algorithm, to represent features better. References to color refer to the online version of this figure
Fig. 5 {Detailed comparison:} the odd rows show the recovered images with original sizes, while the even rows enlarge their highlighted regions. From left to right: results obtained by the algorithms proposed by Adams (2011), Xie et al. (2014) and this study, and the original raster images. References to color refer to the online version of this figure
Image Resolution SR(%) Time(s)
Lena512×51212.03
23.66
36.23
Fruits512×48011.75
23.10
35.19
Flower400×30010.66
21.02
31.61
Falcon400×28010.59
20.86
31.33
Car400×30010.64
21.04
31.61
Fish320×40010.66
21.05
31.66
Leaf373×40010.83
21.36
32.20
Swan286×40010.60
20.95
31.48
Frog357×40010.90
21.44
Table 2 Runtime of our algorithm
Fig. 6 A failure case: (a) is the result of swan by our algorithm at a 1% sampling rate; (b) is the original image. References to color refer to the online version of this figure
[1]   Adams MD . A flexible content-adaptive meshgeneration strategy for image representation. IEEE Trans. Image Process. 2011, 20(9): 2414-2427 doi: 10.1109/TIP.2011.2128336
doi: 10.1109/TIP.2011.2128336 pmid: 21421439
[2]   Demaret L , Iske A . Advances in digital image compression by adaptive thinning. Ann. MCFA. 2004, 3: 105 -109
[3]   Demaret L , Dyn N , Iske A . Image compression by linear splines over adaptive triangulations. Signal Process. 2006, 86(7): 1604 -1616 doi: 10.1016/j.sigpro.2005.09.003
doi: 10.1016/j.sigpro.2005.09.003
[4]   Hu SM , Zhang FL , Wang M . et al. . PatchNet: a patch-based image representation for interactive librarydriven image editing. ACM Trans. Graph. 2013, 32(6): 196 doi: 10.1145/2508363.250838
doi: 10.1145/2508363.250838
[5]   Huynh-Thu Q , Ghanbari M . Scope of validity of PSNR in image/video quality assessment. Electron. Lett. 2008, 44 (13) : 800 -801 doi: 10.1049/el:20080522
doi: 10.1049/el:20080522
[6]   Lai YK , Hu SM , Martin RR . Automatic and topology-preserving gradient mesh generation for image vectorization. ACM Trans. Graph. 2009, 28 (3): 85 doi: 10.1145/1531326.1531391
doi: 10.1145/1531326.1531391
[7]   Lecot G , Levy B . Ardeco: automatic region detection and conversion. 2006 17th Eurographics Symp. on Rendering, p. 349: 360 doi: 10.2312/EGWR/EGSR06/349-360
doi: 10.2312/EGWR/EGSR06/349-360
[8]   Liao ZC , Hoppe H , Forsyth D . et al. . A subdivision-based representation for vector image editing. IEEE Trans. Vis. Comput. Graph. 2012, 18 (11): 1858-1867 doi: 10.1109/TVCG.2012.76
doi: 10.1109/TVCG.2012.76 pmid: 22392717
[9]   Liu DC , Nocedal J . On the limited memory BFGS method for large-scale optimization. Math. Program. 1989, 45 (3): 503-528 doi: 10.1007/BF01589116
doi: 10.1007/BF01589116
[10]   Sieger D , Botsch M . Design, implementation, and evaluation of the surface_mesh data structure. 2012, Proc. 20th Int. Meshing Roundtable: p.533-550 doi: 10.1007/978-3-642-24734-7_29
doi: 10.1007/978-3-642-24734-7_29
[11]   Sun J , Liang L , Wen F . et al. . Image vectorization using optimized gradient meshes. ACM Trans. Graph. 2007, 26 (3): 11 doi: 10.1145/1239451.1239462
doi: 10.1145/1239451.1239462
[12]   Swaminarayan S Prasad L . Rapid automated polygonal image decomposition. 2006 35th IEEE Applied Imagery and Pattern Recognition Workshop, p.28-33 doi: 10.1109/AIPR.2006.30
doi: 10.1109/AIPR.2006.30
[13]   Xia T , Liao BB , Yu YZ . Patch-based image vectorization with automatic curvilinear feature alignment. ACM Trans. Graph. 2009, 28 (5): 115 doi: 10.1145/1618452.1618461
doi: 10.1145/1618452.1618461
[14]   Xie H , Tong RF , Zhang Y . Image meshing via alternative optimization. J. Comput. Inform. Syst. 2014, 10 (19): 8209-8217 doi: 10.12733/jcis11723
doi: 10.12733/jcis11723
[15]   Xiong SY , Zhang JY , Zheng JM . et al. . Robust surface reconstruction via dictionary learning. ACM Trans. Graph. 2014, 33(6) doi: 10.1145/2661229.2661263
doi: 10.1145/2661229.2661263
[16]   Xu L , Lu CW , Xu Y . et al. . Image smoothing via L0 gradient minimization. ACM Trans. Graph. 2011, 30 (6): 174 doi: 10.1145/2024156.2024208
doi: 10.1145/2024156.2024208
No related articles found!