Accurate Image Registration And The Application In Precision Electronic Assembly

Image registration is the process of finding correspondence between all points in twoimages of a scene. This process enables spatially aligning the images and makes it possible todetect differences between images, fuse information in images, estimate3-D structure of ascene, and locate objects of interest in an image. Registration of images with complexnon-linear, locally dependent geometric distortions, multimodality and high dimensionsbelong to the most challenging tasks at this moment. When registering images with non-linear,local distortions, we are facing with two basic problems-how to match control points andwhat mapping functions to use for registration. Although the second one can be solved at leaston theoretically level by using appropriate radial basis functions, the first problem is generallyunsolvable due to its nature. Since the deformation between images could be arbitrary, we cannot use any definitely approaches to determine point pairs. Another conceptual problem ishow we can distinguish between image deformation and real changes of the scene, that's tosay, how we can determine accurate evaluations of image registration. In this thesis, weproposed a new method of adaptive image registration via Hierarchical Voronoi Subdivisionto handle this difficult problem. The method proposed here can withstand considerableoutliers and nonlinearity between images and yet remain efficient. Various steps are built intothe algorithm to detect and remove incorrect and inaccurate correspondences. Furthermore, anew error measurement of image registration is proposed to reduce the influences of intensitydifferences between reference images and test images.High accurate image registration is needed in the fields of target localization of computervision,3-D reconstruction, remote sensing, and medical systems. During the last decades,large number of researchers had been concentrating on the problem of improving the accuracyof image registration. In this thesis, we propose a high accurate subpixel image registrationalgorithm based on phase correlation and image upsampling. The two-stage coarse-to-finealgorithm for image registration is presented. Firstly, the coarse step uses conventional phasecorrelation method to identify the pixel level shift, and then the fine step uses the matrixmultiply discrete Fourier transform to calculate the upsampled region around the coarse pointto locate the subpixel level shift. Meanwhile, the proof of the equivalence between matrix-multiply DFT and zero padding upsampling FFT is given. Experimental results showthat the proposed method provides performance improvements over conventionalcross-correlation based and phase-correlation based subpixel registration under accuracy,efficiency and noise levels.In the fields of large-scale Integrated Circuit Manufacturing, automatic quality controland target localization in industry applications, especially in the case of precision electronicassembly, fast and high accurate image registration is needed. The benefits of automatic visualinspection are well known in the manufacturing and assembly of circuit boards. So manyadvances have been made in the inspection of bare circuit boards and solder joints. However,relatively few works have been demonstrated for mounting components and improvingplacement accuracy. In this thesis, we have proposed several effective methods of fast andaccurate image registration for the need of electronic applications. These methods are mostlybased on shape descriptor, feature matching, and frequent information or object optimization.Experimental results demonstrate the effectiveness and distinctive features of the methods.Although these methods are developed for the need of precise electronic assembly, they couldbe applied in other fields as well.