修改和諧密集連接網路做電子元件X-ray影像的瑕疵分割;Defect segmentation for X-ray images of electronic components using modified harmonic DenseNet

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/93252

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/93252

Title:	修改和諧密集連接網路做電子元件X-ray影像的瑕疵分割;Defect segmentation for X-ray images of electronic components using modified harmonic DenseNet
Authors:	蔡翔宇;Tsai, Shiang-Yu
Contributors:	資訊工程學系
Keywords:	密集連接網路;和諧密集連接網路;語意分割;DenseNet;HarDNet;semantic segmentation
Date:	2023-07-25
Issue Date:	2024-09-19 16:50:46 (UTC+8)
Publisher:	國立中央大學
Abstract:	電子元件 (electroniccomponents) 是所有電子產品的基本元件，電子元件的品質深深影響所有電子產品的品質;因此控制好電子元件焊在印刷電路板上的品質是目前相關業者重要的議題之一。任何產品的製造總避免不了異常情形產生，因此檢測出印刷電路板上的電子元件瑕疵是控制好 “上件印刷電路板” (printed circuit board assembly, PCBA) 出場品質的重要課題。近些年來，深度學習 (deep learning) 技術的發展突飛猛進，在各行各業都有傑出的表現。自動光學檢測 (automated optical inspection, AOI) 和自動視覺檢測 (automatedvisualinspection,AVI) 領域也不例外，大量引入深度學習技術以同時提升產品的瑕疵檢出率 (detection rate) 與篩除率 (screening rate)。用於檢測印刷電路板之電子元件瑕疵的深度學習技術有辨識 (recognition)、偵測 (detection)、分割 (segmentation)、異常偵測 (anomaly detection)、等。在本論文的研究中，我們將以語意分割 (semantic segmentation) 技術來找出印刷電路板之電子元件的瑕疵區塊並分類。我們修改了和諧密集連接網路 (HarDNet-MSEG) 做電子元件 X-ray 影像的瑕疵分割與分類;修改內容包括: i.將解碼器架構設計成 UNet++的形式，使用五層解析度的特徵圖，利於穩定地找尋小瑕疵或較準確的邊界; ii.將感受視野區塊做更改，當中減少複雜的卷積，有利於捕捉小範圍特徵; iii.在最深層的編碼器中加上注意力模組，讓網路排除不必要的刺激，能更關注於重要特徵。在實驗中，我們收集了 979 張電子元件的瑕疵 X-ray 影像，將其分為訓練集有 881 張及測試集 98 張，在訓練時會將訓練資料擴增八倍。原本 HarDNet-MSEG 的訓練集 MIoU 為 91.53%，召回率為 95.21%，精密度為 95.69%，測試集的 MIoU 為 78.23%，召回率為 83.67%，精密度為 92.05%;經過本研究修改後，訓練集 MIoU 為 95.27%，召回率為 97.86%，精密度為 97.83%，測試集的 MIoU 為 86.56%，召回率為 92.59%，精密度為 93.95%。;Electronic components are the fundamental elements of all electronic products. The quality of electronic components deeply affects the quality of electronic products. Therefore, keeping the quality of electronic components soldered on printed circuit boards is one of the important issues for the relevant industry tasks. Any manufacturing process inevitably encounters abnormal situations, so detecting defects in electronic components on printed circuit boards is a crucial aspect in ensuring the outgoing quality of printed circuit board assemblies (PCBAs). In recent years, there is a remarkable advancement in the development of deep learning techniques, they demonstrated outstanding performance in various industries. The fields of automated optical inspection (AOI) and automated visual inspection (AVI) also energetically engage the technique to simultaneously improve the defect detection rate and screening rate of products. Deep learning techniques have been used for detecting electronic component defects on printed circuit boards include recognition, detection, segmentation, anomaly detection, etc. In this studying, we focus on the application of semantic segmentation technique to identify and classify the defective regions of electronic components on printed circuit boards. We modified the Harmonic DenseNet MSEG (HarDNet-MSEG) for defect segmentation and classification in X-ray images of electronic components. The properties of this studying include: i. The decoder architecture was designed in the form of UNet++. It utilized four levels of resolution feature maps, which facilitated stable detection of small defects and more accurate boundaries. ii. The Receptive Field Blocks (RFBs) module was modified by reducing complex convolutions. This modification was beneficial for capturing small-scale features effectively. iii. An attention module was added to the deepest layer of the encoder. This allows the network to eliminate unnecessary stimuli and focus more on important features. In the experiments, we collected 979 X-ray images of electronic components with defects. These images were divided into a training set of 881 images and a test set of 98 images. During training, sample data were augmented into eightfold. The original HarDNet-MSEG model achieves the performance on the training set listing as mean intersection over union (MIoU) of 91.53%, recall of 95.21%, and precision of 95.69%. On the test set, it achieves MIoU of 78.23%, recall of 83.67%, and precision of 92.05%. After the proposed modification, the modified model′s performance is remarkably improved. On the training set, it achieved a MIoU of 95.27%, recall of 97.86%, and precision of 97.83%. On the test set, it achieved a MIoU of 86.56%, recall of 92.59%, and precision of 93.95%.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	23	View/Open

社群 sharing

Loading...