| The fully automatic gold wire bonder is a modem micro-electronic wrapping equipment that utilizes interdisciplinary technologies such as fine mechanics, auto control, image recognition, optics, ultrasonic heat-pressing sealing and so on. It is mainly applied to bond the chip pad with the lead-frame in the post-wrapper process of integrate circuit manufacture.Works that should be done here is to develop the image recognition software for the fully automatic gold wire bonder, and the following researches and experiments are conducted.First, the working principle of the fully automatic gold wire bonder is introduced, two schemes that can be used, that is, the image computer processing system and the software image processing system, are analyzed, and the latter one is adopted according to the demands of our research.Second, some digital image pre-processing techniques are provided, and two image enhancing techniques, the gray histogram trimming and the median filter are discussed. Contrast experiments are conducted on real chip images to check effects of five usually used median filter operator, that is Roberts operator, Sobel operator, Prewitt operator, Krisch operator and Laplacian operator. Experimental results show that the Roberts operator is to our choice.Third, some commonly used fast image matching algorithm based on gray matching are discussed, including FFT algorithm, Multitude Sorting Relational algorithm, Sequential Similarity Detection algorithm, Multiresolution Pagoda Structure algorithm and Genetic algorithm. Their strong and weak points are pointed out by experiments and discipline analysis. It is demonstrated that all those algorithms do not satisfy the real time demand (<25ms/line) of the image recognition system for the fully automatic gold wire bonder, and the first four algorithms can't meet the matching precision demand of the chip image that has a rotating degree from-7 to 7 too.Next, two algorithms that meet the real time demand of the system and one algorithm that have application value are proposed, that is TS algorithm, TS-GA algorithm and MPSA-GA algorithm.TS algorithm is a fast algorithm based on two dimensions template sampling. In the algorithm the search step and sample sequence are set to reduce the computationalcomplexity. The algorithm is broken into two search parts, which settles down the problem of target subimage rotating and guarantee to find the matching positions precisely. Furthermore, SSDA algorithm is utilized to reduce the computational complexity of the algorithm. The simplicity of the algorithm facilitates its hardware implementation. Experiments show that the real time demand (<20ms) and precisiondemand (?um ) of the system are met well and the algorithm is robust and iinmuneto noise.Besides the preprocessing of the template image like TS, TS-GA exploits the global searching mechanism of the genetic algorithm to fasten the convergence of the algorithm to the sub-optimal solution to the problem. On this basis, a precisely locating phase is introduced to fix the location and rotating degree of the optimal target sub image in the scene image. Experiments show that the algorithm also satisfies the demand of the system. In fact, TS-GA is slower than TS due to the time-consuming precision matching step. However, the algorithm can find the target for a second time when the first searching step missed the target due to its non-deterministic searching strategy. As such, this algorithm is chosen to be the alternative to TS when it misses the target.MPSA-GA is based on MPSA and GA, experimental results demonstrate that the algorithm can reduce matching time to some degree. As the first algorithm we thought out, it has elucidating influence on the above two algorithms. Here, we also give a brief introduction to it.At last, the generalization and the anti-noise ability of these three algorithms are tested using an ordinary image, and the experimental results show that these algorithms meet those demands well; therefore, those algorithms can... |
PDF Full Text Request