Research On Optical Inspection Algorithms Of The Defect On Surface Of The Container

This dissertation firstly applies optical inspection technology to defect detection of surface of the container, and expand the application field of this technology. Automated optical inspection is deeply studied, including algorithms of image enhancement, image segmentation, multiresolution analysis based on wavelet, phase measurement profilometry, calibration of measurement system and registration for 3-D point cloud, etc. The main contents of this dissertation are as follows:2-D image processing is studied to analyze the gray characteristics for the problem of hole detection on surface of the container. An improved spatial low-pass filter with an undetermined parameter is proposed, and an unsupervised method of automatic parameter selection for image enhancement is proposed based on Otsu's method. This algorithm combines image enhancement with image segmentation as one procedure through a modified discriminant criterion. The optimal parameter of the filter is selected by the discriminant criterion given to maximize the separability between object and background. The optimal threshold for image segmentation is computed simultaneously.Multiresolution analysis is studied to analyze the texture and color characteristics on metal surface for the problem of corrosion detection on surface of the container. The image is divided into non-overlapping blocks of size 8×8 pixels. For each block, db6 wavelet is applied to decompose it into subimages in various scale and wavelet spaces. Feature vector is constructed to describe the texture and hue information according to the energy in each block. K-mean classifier is trained using prior knowledge, and blocks are classified into corrosion and corrosion free classes based on nearest neighbor classifier.To detect suface deformation of the container, 3-D measurement based on active structured light is applied. This dissertation studies the working mechanism of phase measurement profilometry, and introduces hardware constitution of adopted measurement system. Algorithms for phase unwrapping are deeply studied, and two factors impacting the accuracy are discussed, i.e. width selection of digital sine-wave grating projected and detection of residues. It is noted that half-cycle width of projected fringe pattern must cover maximum sudden change of height on surface of the object, and the residues must be romoved totally. This dissertation analyzes the four constructions of residues in detail, firstly introduces the technique of pattern classification into noise-residue elimination, and designs a feasible filter to remove noise-induced inconsistencies. On condition of much noise, all the residues can be filtered effectively, and the wrapped phase map can be considered as reliable input signal to the algorithm for path-dependent phase unwrapping. Consequently, the unwrapped phase map is almost correct, not just consistent.Given the unwrapped phase map computed based on phase measurement profilometry, the correspondence between unwrapped phase and height value is studied. Relations of coordinate transform among camera coordinate system, projector coordinate system and world coordinate system are derived based on perspective projection models for camera and projector. Calibration model of measurement system is derived with uncollimated illumination according to idea from Du[142,143], and eleven undetermined parameters in the model are determined using two gauge objects with known heights and nonlinear optimization algorithm of least-squares. Besides, the reference plane and the gauge objects are used for the system calibration only, and are not required in actual 3-D shape measurement.Registration algorithms for 3-D point cloud are investigated, given the unwrapped phase map computed based on phase measurement profilometry. The working mechanism of classical ICP algorithm is deeply studied, and its advantages and limits are analyzed. Registration based on angular-invariant is then proposed. The relationship of coordinate transform between data set of model and measured data set is calculated accurately using three pairs of nearest neighbor points without setting accurate initial estimates. Because the proposed angular feature is invariant to rotation, translation and scale transformation, the registration result will converge to global minimization rather than local one. Simulation experiments verify the validity of the proposed registration algorithm based on angular-invariant feature: the proposed algorithm is more accurate than ICP algorithm, and faster than ICP and the algorithm based on curvature-invariant feature.