- 學位論文
閻推薦: X光影像印刷電路板瑕疵檢測參數推薦系統
YenRecommend: Parameter Recommendation for Printed Circuit Board Defect Inspection System for X-ray Images
摘要
本篇論文提出了對於瑕疵檢測機台的檢測過程中，不同特性組成的印刷電路板影像，所對應的最適當機台檢測參數組，藉此達到良好的檢測品質，並且同時降低檢測時間成本。因為在機台做瑕疵檢測的過程當中，現場應用工程師 (FAE: Field Application Engineer)，也就是客戶服務工程師會對客戶送進來的一批新的印刷電路板，根據自身經驗作分析並給出一個優良的參數組合讓以供機台檢測；然而每一輪的機台檢測結果都需要進一步調整，一次次的達到最合適的參數組，這個過程不僅費時費財，也消耗現場應用工程師的人力成本，因此我們提出一套系統做參數推薦，藉以替代現場應用工程師在前半段一步步調整參數組的動作，以期達到參數推薦自動化的效果，希望在新機台或新印刷電路板初次檢測時能快速縮短起始時間，找出優良檢測參數，達成優良檢測成果，也就是低漏檢率與低假警報。近年來，強化學習 (Reinforcement Learning)在機器學習與深度學習上都有蓬勃的發展，我們從影像抽取方向性與旋轉特徵 (ORB: Oriented FAST (Features from Accelerated Segment Test) and Rotated BRIEF (Binary Robust Independent Elementary Features))並分群 (Clustering)，並根據每一群去做參數組初始值的設定；接著加上機器學習去讓我們的系統了解到每一群在每一個情境下最合適的下一步(如何調整各個參數)。在一組新的影像輸入後，我們的系統會找尋最符合這組影像的群，接著給予最佳的初始參數組以及後續的調整參數步驟，達到參數推薦自動化的效果。
並列摘要
In this thesis, we propose a parameter recommendation system for defect inspection machine due to different characteristics of printed circuit board images, to reach excellent inspection quality and also reduce initial time cost. During defect inspection process, FAE (Field Application Engineer) must search and configure excellent parameter set for a batch of new printed circuit boards according to their own experiences. However, each result requires further adjustment to reach the most appropriate parameter set step by step. This procedure is not only time-consuming but also wasting labor costs. Therefore, we propose parameter recommendation system to accelerate those adjustments by FAE initially, so we can recommend parameter set automatically. Recently, machine learning flourishes. We extract image features to be inspected by ORB, cluster on those features, and set up the initial values for the parameter set. Next, we apply machine learning methods to let our system understand which is the best choice for next step on each cluster in every situation. When a batch of new printed circuit board images arrive, our system will find the most suitable cluster for them, give them good initial values as well as the following adjusting steps, to achieve automatically recommending parameter set.