图像匹配结合SIFT和区域匹配程序2012国际会议上控制工.pdf

Image matching combine SIFT with regional SSDAQiu Wentao1,2,Zhao Jian1*,Liu Jie1(1 Changchun Institute of Optics,Fine Mechanics and Physics,Chinese Academy of Sciences,Changchun 130033,China,E-mail:;2 Graduate University of the Chinese Academy of Sciences,Beijing 100039,China)AbstractImage matching is at the base of many computer vision problems,such as object recognition or image stitching.Standard SIFT provides poor performance when images under viewpoint change conditions and with similar corners.Hence,we propose a matching algorithm combine regional SSDA with simplified SIFT algorithm.We demonstrate through experiments that our algorithm yields better performance in images ofviewpoint change and similar feature points.Besides,the simplified algorithm cut down about half the time was originally needed in our tested images.Keywords-Image matching?SIFT?SSDAI.INTRODUCTION Image matching is an essential part of many modern computer vision systems that solve tasks such as object or scene recognition,stereo correspondence,image indexing and it is the basis of image stitching,image fusion 1,2.It is a technique that wildly used in processing medical images,remote sensing images etc.Image registration is to find a transformation between two or two more images,which is of the same scene obtained from different time or different sensors or different viewpoint,enables their identical in spatial locations.In literature 1,2,methods of image matching can be broadly divided into two categories:image matching based on gray intensity and based on image feature.Image matching based on intensity usually make use of the intensity information to build a similarity function between two images,such as sequential similarity detection algorithm(SSDA)3,phrase correlation and Fourier-Mellin transformation etc.While image matching based on feature oftentimes extractfeatures such as corners,contours,interlines,curvature and synthesized descriptors etc.Since it doesnt rely on image intensity thus reducing the probability of disruption by noise,it became the most popular methodology that used currently.Scale-invariant feature transform(SIFT)4,5,6 is invariant to image scaling and rotation,and partially invariant to change in illumination and 3D camera viewpoint.In addition,large number of features detected is of highly distinctive.All these properties make SIFT the most popularly studied recently.Most related research efforts focused on designing and learning effective descriptors or on speeding up the computation in detection components,incorporating the applicable criteria that follow.II.REVIEW OF SIFT ALGORITHMStandard SIFT,as described in 4,consists of five major stages of computation to detect and match feature points,which is shown in Fig.1.Figure 1 SIFT computational process?Scale-space 7 extrema detection,extremum detection in both scale space and image coordinate plane.A Gaussian pyramid is constructed and candidate points are extracted by scanning local extremum in a series of DoG(Difference of Gaussian)images.?Keypoint location,candidate points are localized to sub-pixel accuracy through fitting three-dimension quadratic function,and unstable points of low contrast or strong edge response are eliminated.?Orientation assignment,orientations are assigned to each keypoint location based on statistical result of local image gradient directions.?Keypoint descriptor,SIFT descriptor is the representation of statistical local image gradients which measured at the selected scale in the region around each keypoint.?Matching the SIFT feature points.Calculate Euclidean distance between each SIFT descriptors to match points.In later estimation parameters of transformation,use RANSAC 8 to eliminate outliers to ensure the accuracy of parameters estimation.III.SIFT COMBINE WITH REGIONAL SSDABased on our observations,when it comes to images that with change in viewpoint and with similar corners within images,SIFT cannot get a satisfying result.In order to tackle this problem,we propose a simplified SIFT combine with regional SSDA.In this section,we elaborate our proposed algorithm,using two pair of images with size of 816?616.There are rotation,real-world viewpoint change,illumination and scale change between reference image and test image.Our proposed algorithm run on Intel E5200 Pentium(R)Dual-CoreCPU 2.5GHz PC.We first apply SIFT to these images,and the results show in Fig.2 and Fig.3.We can see only one correct match in Fig.1 and no correct match obtained in Fig.2.Figure 2 SIFT matching result of first pair imagesFunded by Key Project of Science and Technology Department of Jilin Province(Grant No.20100310)2012 International Conference on Control Engineering and Communication Technology978-0-7695-4881-4/12$26.00 2012 IEEEDOI 10.1109/ICCECT.2012.781772012 International Conference on Control Engineering and Communication Technology978-0-7695-4881-4/12$26.00 2012 IEEEDOI 10.1109/ICCECT.2012.78177Figure 3 SIFT matching result of second pair imagesA.regional SSDAIn order to enhance the ability of SIFT to distinguish similar corner within image,we came up with pre-matching different regions of images.To obtain this,we first apply edge detection to images.Here we use sobel operator to get edges due to its effectiveness.According to the connectivity and content of edges in one region i.e.we scanning rows and columns,if there is sharp change of content and connectivity in images,we set appropriate threshold to divide images into different parts.Later,SSDA was calculated between different parts to pre-match different regions.Consequently,if a region has high score from SSDA and similar edge content,then we assume these two regions were pre-matched.Since not all different edge regions are of the same size,in this process,down sampling were used to confined region to the same size to apply SSDA.By pre-match image regions,we confined our match in similar regions to avoid mismatch between analogousfeature points in other region.As is shown in Fig.4 and 5,the red rectangle area are pre-match similar region obtained from this stage in edge form images extracted by sobel operators.Figure 4 pre-match regions of first pair imagesFigure 5 pre-match regions of second pair imagesB.simplified SIFTSince we get pre-match areas in both images,only these matched regions were participated in the following SIFT process.In order to further more reduce the computational burden,one octave simplified SIFT algorithm was proposed.Reducing octaves in SIFT will decline the coverage of different scales,we compensate this drawback by increasing the value of k which is the ratios of scale between intervals.Nevertheless,the number of matching points decreases with only one octave images.By experiments,we found that bynormalizing the scale in DOG images provides more matching points.Here,we normalize scale in DOG images by the following equation 9.DOG=(DOG-min(DOG)/(max(DOG)-min(DOG).As a result,we obtain an increase in number of matching points while lower the computation time,thus improving the stability of matching algorithm.Due to combining with pre-match result in matching the SIFT feature points stage,we confined our calculation only in similar regions.Experimental results of tested images are shown in Fig.6 and Fig.7.From Fig.6,we can see that 9 pairs of correct matching points were obtained,and in Fig.7,we get 6 pairs of correct matching points.Applying our simplified improved SIFT,a considerable amount of correct matching points were attained.Figure 6 proposed algorithm matching result of first pair imagesFigure 7 proposed algorithm matching result of second pair images178178IV.DISCUSSION OF EXPERIMENTAL RESULTSBy contrast in table 1,our proposed algorithm acquire more correct matching points,as is shown in Fig.2,Fig.3 and Fig.6,Fig.7.Some factors may account for this improvement.For one,by pre-match the image pairs lessen the possibility of mismatching similar features points through confining matching calculation within matched regions.For another,normalizing the response in DOG images heighten the comparability of different interval response in DOG images,thus increasing the number of matching points.From table 2,in the tested image matching experiments,our method decline about half the time it was originally consumed.The reason for that is quite self-explanatory.Table 1 comparison correct numbers of matching pointsStandard SIFTProposed algorithmFirst pair images19Second pair images06Table 2 time consuming comparisonStandard SIFTProposed algorithmFirst pair images 921ms486msSecond pair images1687ms754msV.CONCLUSIONAiming to solve registration of images with change of viewpoint and similar features within image,we propose a simplified SIFT combine with regional SSDA.Not only can our method extend the ability to match images under change of viewpoint,but it also cut down the computational burden.Though there are issues remain.Future work will focus on generalize the pre-match method to different images.ACKNOWLEDGMENT This research was supported in part by Key Project of Science and Technology Department of Jilin Province(Grant No.20100310).REFERENCES1Tuyte T,Mikolajczyk.Local invariant feature detectors:A survey.Foundations and trends in computer graphics and vision 2008,v3.2Mikolajczyk K.Schmid C.A performance evaluation of local descriptors J.IEEE Transactions on Pattern Analysis and Machine Intelligence,2005(27):1615-1630.3MI Chang-wei,LIU Xiao-li,XU Ming-you;An Advanced Algorithm Based on SSDAJ;Journal of Projectiles Rockets.Missiles and Guidance;2004-014David G.Lowe.Distinctive image features from scale-invariant keypoints J.International Journal of computer vision 2004,60(2),91-110.5Rob Hess.An Open-Source SIFT Library.Proceedings of the international conference on multimedia,2010.6OpenCV.http:/ T.Scale-space for discrete signals J.IEEE Transaction PAMI,1980,207:187-207.8M.A.Fischler and R.C.Bolles.Random sample consensus:a paradigm for model fitting with applications to image analysis and automated cartography.Communications of the ACM,24(6),1981.9Di Nan,Li Guiju,Wei Yajuan.Image matching technology based on SIFT for terminal-guiding system.J;Infrared and laser engineering 2011,40(8)179179