| 摘要: | 隨著科技日新月異，科技的急速進步驅動硬體技術的飛躍發展，這對電路板的需求不斷攀升，同時也提高了對品質的嚴格要求。深度學習技術因其卓越的應用潛力而備受矚目，不僅在工業界，也在日常生活中發揮了關鍵作用。舉例來說，交通管理領域使用深度學習技術，實現路口科技執法系統，它可以自動偵測紅燈違規或超速行駛，提升道路的安全，同時也提高執法的效率。
目前在電路板印刷相關產業中，印刷電路板（PCB）的良率檢測主要依賴自動光學檢測（AOI）系統和人工檢測。然而AOI系統常常出現缺陷判斷誤差，這導致需要大量人力介入，從而增加了生產成本。為了有效降低PCB檢測的人力成本，本研究提出了一種基於深度學習的檢測技術，用於辨識PCB上的缺陷。我們的目標是建立一個深度學習模型，以高度精確地過濾掉AOI系統標記的「偽缺陷」，從而提升檢測的準確性和效率。
本研究經過一系列嚴謹的測試與評估後，選擇YOLO神經網路作為模型訓練的主架構。近年來YOLO因其在物件偵測領域的卓越性能，已在學術及工業界廣泛應用。本研究將瑕疵視作特定物件，透過深度學習進行細緻的訓練，系統得以高精度地識別並標注瑕疵位置。而模型訓練所用的資料集，則是由合作廠商提供目前AOI系統於產線上所蒐集的瑕疵資料，其中包含了AOI系統錯誤識別的八類瑕疵以及非瑕疵影像資料。
;With the rapid advancement of technology driving the leap forward in hardware techniques, there is an escalating demand for circuit boards, paralleled by increasingly stringent quality requirements. Deep learning technology, recognized for its exceptional potential in applications, plays a pivotal role not only in the industrial sector but also in daily life. For instance, in the field of traffic management, deep learning has been implemented to enable intelligent traffic law enforcement, including technological systems at intersections that automatically detect red light violations or speeding, thereby enhancing road safety and enforcement efficiency.
Currently, in the printed circuit board (PCB) manufacturing industry, the inspection of PCB yield primarily relies on Automated Optical Inspection (AOI) systems and manual checking. However, the AOI systems frequently encounter defect judgment errors, leading to substantial human intervention and thus, increasing production costs. To effectively reduce the labor costs associated with PCB inspection, this study proposes a deep learning-based detection technique to identify defects on PCBs. Our goal is to establish a deep learning model that can accurately filter out the ′pseudo defects′ marked by the AOI systems, thereby increasing the precision and efficiency of inspections.
After a series of rigorous tests and evaluations, this research has chosen the YOLO neural network as the principal framework for model training. YOLO, widely applied in academia and industry for its superior object detection capabilities in recent years, is utilized in this study to treat defects as specific objects. Through meticulous training with deep learning, the system is capable of identifying and marking defect locations with high accuracy. The dataset used for model training is comprised of defect data currently collected by the AOI systems on the production line, provided by our industry partners, including eight types of defects and non-defect image data erroneously identified by the AOI systems. |
