| Summary: | 碩士 === 國立高雄應用科技大學 === 光電與通訊研究所 === 96 === The thesis contains two subjects：wafer sawing-lane inspection and contour detection on wild animals.
Today, there is no doubt that diamond wafer sawing is a viable method for the die separation of microelectronic substrates. However, reality dictates that much careful planning and control over many variables is necessary in order to create the required efficient sawing system. Material type, depth of cut, desired throughput, feed rates, spindle speed, and blade flange design, are but a partial list of the variable components affecting the sawing operation. Also, because of the mechanical constraints induced by sawing, chips corners tend to break easily thus rendering the pieces unusable. This means a large amount of scrap wafers are required for qualifications.
In this thesis, a wafer sawing lane inspector has been designed to suit in a production line environment, and help NXP Semiconductor Co. at Kaohsiung check if the lane width is manufactured to specification. We present an efficient algorithm for sawing lane extraction based on modified YCbCr color processing, 10 × 10 crossed template matching, energy transformation and analysis, and sawing lane localization via local projection along horizontal and vertical directions, respectively. Then, we trace the contour of sawing lane based on 8-adjacency scheme, marks on chips that need to be checked are achieved using distances among sawing boundaries and die walls. Due to the possible translation and rotation errors caused by the x-y table of the mechanism or improper sawing, the orientation of lane calculated by the least squares regression algorithm is carried out to prevent unsatisfactory positioning from reaching the width inspection. 30 images were used in our test dataset. Experimental results show that both the accurate acceptance rate and the accurate rejection rate are 100%, while both the false acceptance rate and the false rejection rate are 0. The running system is AMD Athlon 64 2.0 GHz CPU, which spends 0.72 second in average on processing one 640 × 480 image.
Concerning about the contour detection of wild animals, we try to separate the target objects from the complicated backgrounds. We analyze the energy spectrum of the target object and perform the connected component labeling concomitantly.
|
|---|
