Research On Key Technique Of Wire Bonding Vision Inspection

Computer vision inspection technique is the core technology of high speed and high precision pattern recognition system of wire bonder. When the imaging equipment is under certainty condition, the vision inspection algorithms determine the speed and accuracy of the vision inspection system, further influence the cost and reliability of microelectronic products. In recent years, with the development of microelectronic products, especially the development of micro-system technology, the dimensions of IC chips and MEMS products have been scaled down and the structures of which have been increasingly complex. There is a higher demand on image processing, vision positioning and defect inspection algorithms, which have become one of the key factors for higher speed and accuracy flexible wire bonding. The vision inspection algorithms for wire bonding are studied in depth, including the introduction of research developments home and abroad, and presentation of wavelet cellular automata based de-noising algorithm, multi-feature fusion based positioning algorithms and Curvelet moment feature based defect inspection algorithm.The performance of image de-noising algorithms affects the accuracy of subsequent feature extraction, matching location and defect inspection. The improved fast adaptive weighted median filtering algorithm and bilateral filtering algorithm are chosen to eliminate the impulse noise and Gaussian noise in the vision inspection system respectively. In order to remove the mixed noise of the system effectively, a wavelet cellular automata de-noising algorithm is presented according to the different characteristics of noise and signal in the frequency domain and spatial domain. Direction information measure and edge order measure of the pixels are selected as the evolution rules and the accurate noise information is obtained by automatic evolution of cellular automata. The simulation results show that the proposed algorithm effectively solve the edge blur problem of the existing algorithms with an increase in PSNR, the mixed noise is efficiently eliminated while edges and detail information of the image is preserved well.Aiming at the matching location in vision inspection, two feature extraction algorithms are presented and on the basis of which two matching location algorithms are realized.The first is the mixed moment invariant based feature extraction algorithm. According to the limited description ability of the single feature, the mixed moment invariant feature is presented which combines the advantages of Zernike orthogonal moments invariant and Wavelet moment invariant, the former can well describe the overall feature of the image while the latter can provide the details in local and nice anti noise performance. The rotation invariant and the anti noise ability of the proposed feature vector are proved experimentally. For the IC chip image, a mixed moment invariant based matching location algorithm is implemented. The coarse-to-fine search strategy is adopted to accelerate the matching and the normalized cross-correlation function is utilized to locate the best matching position. The proposed algorithm fully takes into account the global and local information of the image and has obvious advantages over the idempotent moments based matching algorithms.The second is SIFT-MIC based interest points detect method which is suitable for the MEMS image with rich detail information. By selecting the SIFT feature points with MIC corner properties, the salient feature points are obtained, meanwhile, the computational complexity of the matching algorithm is also decreased effectively. Experimental results demonstrate the rotation, translation invariance and reliability of the SIFT-MIC operator. On the basis of this, random sample consensus (RANSAC) algorithm is employed to eliminate potential false matching points and the translation and rotation parameters are calculated using singular value decomposition. SIFT-MIC operator based matching location algorithm effectively overcomes the shortcoming of the SIFT and improve the positioning accuracy and detection efficiency.To aim at the difficulty of correctly extracting the features of defect regions, the Curvelet moment feature extraction algorithm is put forward utilizing the multi-resolution, time-frequency localization properties, good directivity of Curvelet transform and based on the deep analysis of the sub-bands coefficients. To ensure the universality of the Curvelet moment feature, MPEG-7 contour shape databases are applied to test the invariance and classification performance and the simulation results verify the feasibility and validity of the algorithm. The extraction of defect regions are realized with referential method by the designed preprocessing scheme, the shape characteristics of defect regions are described using the Curvelet moment, and finally the support vector machine(SVM) which suits for the small sample situation is applied to complete the classification of the defects. The experimental results show that the proposed algorithm has the better classification ability compared with the traditional algorithms and has important application value for the recognition of wire bonding product defects.The structure of image acquisition system for wire bonding is given and the work flows of teaching and bonding process are presented based on the operation principle of vision system. This paper also designs and develops testing software based on the proposed matching location algorithms and defect inspection algorithm.To sum up, this paper studies the matching location and defect inspection related technologies for wire bonding in depth, and does experiments and verifies the presented algorithms by using IC and MEMS images, and results show that the algorithms put forward by the paper gain better detection and classification effects. The achievements of the dissertation are of theoretical significance and practical value for the development of computer vision inspection, and will drive the application of computer vision in wire bonding pattern recognition.