- 學位論文
類神經與平行處理於電感缺陷偵測之應用
Inductance Defect Inspection Using Neural Networks and Parallel Processing Techniques
摘要
晶圓代工是目前台灣的主要產業，也是目前台灣的經濟命脈所在，全世界的晶圓代工，台灣就占了將近全世界的70%，為全世界第一，其他的電子資訊產品亦不在話下。 在電子資訊產品中，有幾項是常常需要動用人工去檢查其產品是否有瑕疵的，其中以PCB電子電路板跟電感器的量最多。要抓出這些產品的瑕疵，員工必須經過大量的訓練，而且必須長期的在放大鏡或顯微鏡前與這些商品長時間接觸，久而久之會產生疲勞倦怠，看久了會使的視力和判斷力變差，反應變慢，注意力越來越無法集中的情況下，抓瑕疵的效率也會變得較差，如此一來不良率也會跟著提升。因此，電腦視覺的技術來做缺陷偵測便具有效率穩定的優勢，可以減少人為因素造成的誤判和產能衰弱等等。 本研究使用了水平垂直投影法來擷取樣本的邊界，利用SUSAN角偶偵測來得出樣本的定位點，接著再利用類神經分類(SimNet)來判斷樣本的好壞，最後利用平行處理來加速處理。在本研究的結果裡，1000個樣本中有927個是可以成功的定位而且辨識正確的，其成功率達到92.7%。
並列摘要
The IC foundry is the major industry of Taiwan, and it is also considered the economic lifeline of Taiwan. Taiwan take the mostly 70% of the IC foundry in the world, and it is the No.1 in the world. Many IC foundry products need intensive manpower to inspect the defects. The inductance is considered one of them. The worker needs to train for a long time to have qualified experience for inspecting the product. Moreover, they should take a long time to stay with the microscope or the magnifier during the inspection process. Since people might feel tired after long period of time of inspection, and their vision and judgment might get deteriorated which will further affect the quality of production. The computer vision technology is one way in conquering this problem. In general, computer is stable, accurate, and faster than human being. Many enterprises use the computer vision technology to inspect their product defects, and the result is satisfactory. The objective of this study is to detect the defects of inductance. The first step is to find the inductance border by using the horizontal projection and the vertical projection. Secondly, we use the corner detect technology call “SUSAN” to find the corners. Thirdly, we train and test the inductance samples using the neural network SimNet. Finally, we use parallel processing techniques written in MPI to speedup the inspecting process. There are 927 out of 1000 samples which we can correctly locate the position and performing inspection, the inspection accuracy is 92.7%.