- 學位論文
基於區域卷積神經網路的積體電路板影像辨識與物件偵測模型
Region-based Convolutional Neural Networks-based Image Recognition and Object Detection Model of Integrated Circuit Board
摘要
近年來工業4.0是全球性的熱門議題，在工廠生產的製造過程中，透過整合計算、通訊與控制的虛實化系統，連結人工智慧技術與物聯網來建構智慧製造(Smart Manufacturing)的流程，將成為未來智慧工廠的趨勢。然而，在過去製造業的發展中，彈性製造系統(Flexible Manufacturing System, FMS)是各工廠發展的重點，其中在品質管理中的檢測過程，儘管應用了自動光學檢驗技術(Automated Optical Inspection, AOI)，但其篩選的標準過高，淘汰約30%的產品，還是需要耗費大量的人力與時間執行第二階段檢測，在辨識完整性和人力方面，並不能做到全自動且良好的管控。 在本論文中，將整理並分析目前較為先進且廣泛應用的幾種區域型卷積神經網路模型(R-CNN)，解析各種模型在發展上的問題和方法，並且提出一個適合應用在積體電路板(Integrated Circuit Board, ICB)製造流程與場域中的即時影像辨識模型與架構。本論文第一階段的目標，將蒐集並使用不同種類的積體電路板做為模型訓練資料集，建立初步階段的影像識別模型，讓該模型可以根據不同的特徵點來做分類與預測不同類別的積體電路板；第二階段將探討優化模型方法，在本論文中使用資料增強方法最後的平均準確率可達96.53%，最後針對模型的應用性作探討，以偵測積體電路板之晶片方向性探討，可達98%正確率，且辨識時間不超過1秒可達到智慧製造場域需求之即時性，有助於在智慧製造過程中提供良好的物件影像辨識品質，節省測試人力，並提升整體製造過程的產能與產品良率。
並列摘要
Industry 4.0 has been a hot topic in recent years. In the manufacturing process of the factory, the process of constructing smart manufacturing through the integration of the cyber-physical systems (CPS), artificial intelligence (AI), and Internet of Things (IoT) technology, will become the trend of smart factories in the future. In the past, the development of the manufacturing industry, the flexible manufacturing systems (FMS) is the key of the development of smart factory. However, most parts of the quality management process still need to be implemented by Automated Optical Inspection(AOI) methods. But it method only achieve 70% threshold and need a lot of human resources and time to perform the second phase of testing. In terms of recognition accuracy and integrity, and equipment maintenance, automatic and good monitoring and control is not achieved. For solving the above problems, this project will propose an approach for designing a regional convolutional neural network (R-CNN) image recognition model for smart manufacturing fields. In this study, several R-CNN, which are currently more advanced and widely used, will be collated and analyzed, and then the development problems and solutions of various models will be discussed. Then, this study will propose a real time image recognition model and architecture which suitable for the IC Board manufacturing process. The goal of the first phase of this study is to collect different types of integrated circuit boards and chip parts as training data sets, and establish a preliminary stage image recognition model. The second phase the optimization model method will be discussed. In our research, data augmentation method can reach 96.53% average accuracy. The third phase the applicability of the model is discussed to detect the wafer directivity of the integrated circuit board, which can reach 98% correct rate, and the identification time is less than 1 second to achieve the immediacy of the smart manufacturing. It will help to provide the accurate and instant objects recognition in the smart manufacturing process. The proposed model can increase the yield rate of the production line and the overall equipment effectiveness in the manufacturing process.
參考文獻
延伸閱讀
- Shen, Z. Y. (2017). 結合場景及物件資訊之深度卷積神經網路物件偵測 [master's thesis, National Taiwan University]. Airiti Library. https://doi.org/10.6342/NTU201702383
- 謝承曄（2022）。基於卷積神經網路的多層次醫學影像檢索〔碩士論文，國立政治大學〕。華藝線上圖書館。https://doi.org/10.6814/NCCU202201220
- Lee, K. Y. (2021). 非監督式對比學習卷積神經網路用於多類別共同物體分割 [master's thesis, National Tsing Hua University]. Airiti Library. https://www.airitilibrary.com/Article/Detail?DocID=U0016-2412202116165067
- 廖昱唐（2022）。基於卷積神經網路及影像強化之工地安全帽偵測系統及應用〔碩士論文，國立臺灣大學〕。華藝線上圖書館。https://doi.org/10.6342/NTU202201200
- 陳衍碩（2022）。Design and Implementation of Convolutional Neural Networks based on Multiple Parallel Processing〔碩士論文，中原大學〕。華藝線上圖書館。https://doi.org/10.6840/cycu202201612