本研究為產學合作案。合作廠商為國內、外知名矽晶圓製造廠, 也是台灣碳化矽半導體的龍頭。碳化矽半導體是矽晶圓製造的新一代材料, 其具備高耐壓與高耐溫操作特性, 因此適合用來製造高壓、高溫、抗輻照功率的半導體功率元件。相對傳統矽半導體, 碳化矽半導體的製造較困難, 且生產成本較高。碳化矽晶錠上之缺陷會極大地限制其電力的性能, 因此有效的缺陷檢驗技術不僅有助於進料檢驗、生產過程中的測試檢驗, 甚至還能提供特定缺陷形成的參考訊息, 進而提高碳化矽的產量、性能, 並減少生產浪費的成本。由於碳化矽缺陷現象不易從表面被觀察到, 通常使用 KOH (氫氧化鉀)蝕刻技術, 使表面顯現蝕刻坑形狀, 以判斷缺陷的種類及計算。本研究目的為建立一套有效的碳化矽半導體缺陷自動化檢測和計數方法。

本研究利用產學方提供之碳化矽晶錠切片(晶片)經KOH蝕刻後的影像資料，運用深度學習物件偵測 Mask R-CNN的方法，並經過實驗設計找出最適合的參數組合，提出適用於的碳化矽缺陷自動化檢測的系統。本研究的成果為(1)同時進行檢測與計數三種類型 (BPD, TED, TSD)的缺陷，(2)偵測一張碳化矽影像的平均時間約30秒，(3)缺陷偵測的績效 mAP(Mean Average Precision)達93%。本研究提供產學方一套效率高(準確且快速)、成本低(節省人工目檢成本)的缺陷檢驗機制，著實為半導體產業製程中關鍵的貢獻。

This is a case study on industry–university cooperation. The cooperating manufacturer is a well-known silicon wafer manufacturing plant, which is a big silicon carbide semiconductors enterprise in Taiwan. Silicon carbide is a new generation of a silicon wafer manufacturing material with high temperature working characteristics, and it is suitable for manufacturing semiconductors components. Compared to traditional silicon semiconductors, silicon carbide semiconductors are more difficult to manufacture and their production costs are higher. Moreover, defects in silicon carbide crystal ingots may affect their electrical performance. Therefore, an effective defect inspection method not only is conducive for inspections but also will provide information about the defects, thereby improving the output and performance of silicon carbide and reducing the production waste cost. Since silicon carbide defects are difficult to observe from the surface, the KOH etching technology is used to form etch pits on the surface to determine the type of defects and calculate the defects. Furthermore, this study aims to establish an effective automatic detection and counting method for silicon carbide semiconductor defects.

Herein, the KOH-etched image data of silicon carbide wafers provided by the cooperating manufacturer are applied to the deep learning object detection model Mask R-CNN to determine the most suitable parameter combination through DOE. Thus, an automatic detection system of silicon carbide defects is realized. Importantly, (1) the proposed method can simultaneously detect and count three types of defects (BPD, TED, and TSD), (2) the average detection time for a silicon carbide image is about 30 s, and (3) the mean average precision of the defect detection can reach 93%. Thus, this study provides a high-efficiency and low-cost defect inspection system that saves manual visual inspection costs for cooperating manufacturers, which is a key contribution to the semiconductor industry’s manufacturing process.

摘要 i
英文摘要 ii
致謝 iii
目錄 iv
圖目錄 vi
表目錄 vii
第 1 章 緒論 1
1.1 研究背景 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 研究動機與目的 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 符號與名詞定義 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3.1 符號定義 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3.2 名詞定義 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
第 2 章 文獻探討 4
2.1 深度學習方法 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 物件偵測模型 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2.1 R-CNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2.2 Fast R-CNN . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.3 Faster R-CNN . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.4 Mask R-CNN . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Mask R-CNN 模型應用 . . . . . . . . . . . . . . . . . . . . . . . . . 6
第 3 章 研究步驟與方法 7
3.1 資料蒐集 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.2 資料標記 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.3 深度學習架構 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.4 Mask R-CNN 架構 . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.4.1 Backbone . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.4.2 Neck . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.4.3 Head . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.4.4 Loss Function . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.4.5 績效指標 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.5 最佳參數設計方法 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.5.1 數據增強 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.5.2 超參數調整 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.5.3 實驗設計 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
第 4 章 實驗結果 24
4.1 Mask R-CNN 預設模型 . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.2 2
9−4部分因子實驗 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.3 驗證最適模型 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.4 碳化矽缺陷偵測結果比較 . . . . . . . . . . . . . . . . . . . . . . . . . 28
第 5 章 結論與未來展望 31
5.1 結論 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.2 未來展望 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
第 6 章 參考文獻 33

[1] Linqin Cai et al. “Mask R-CNN-based detection and segmentation for pulmonary nodule 3D visualization diagnosis”. IEEE Access 8 (2020), pp. 44400–44409.
[2] Young-Jin Cha et al. “Autonomous structural visual inspection using region-based deep learning for detecting multiple damage types”. Computer-Aided Civil and Infrastructure Engineering 33.9 (2018), pp. 731–747.
[3] Ross Girshick. “Fast r-cnn”. Proceedings of the IEEE international conference on computer vision. 2015, pp. 1440–1448.
[4] Ross Girshick et al. “Rich feature hierarchies for accurate object detection and semantic segmentation”. Proceedings of the IEEE conference on computer vision and pattern recognition. 2014, pp. 580–587.
[5] Kaiming He et al. “Mask r-cnn”. Proceedings of the IEEE international conference on computer vision. 2017, pp. 2961–2969.
[6] Jeremiah W Johnson. “Adapting mask-rcnn for automatic nucleus segmentation”. arXiv preprint arXiv:1805.00500 (2018).
[7] Hui Lin et al. “Automated defect inspection of LED chip using deep convolutional neural network”. Journal of Intelligent Manufacturing 30.6 (2019), pp. 2525–2534.
[8] Tsung-Yi Lin et al. “Feature pyramid networks for object detection”. Proceedings of the IEEE conference on computer vision and pattern recognition. 2017, pp. 2117–2125.
[9] Pei Ma et al. “Three-dimensional detection and quantification of defects in SiC by optical coherence tomography”. Applied optics 59.6 (2020), pp. 1746–1755.
[10] Shota Monno et al. “Detection of defects on SiC substrate by SEM and classification using deep learning”. International Conference on Intelligent Networking and Collaborative Systems. Springer. 2018, pp. 47–58. 33
[11] Shaoqing Ren et al. “Faster r-cnn: Towards real-time object detection with region proposal networks”. IEEE transactions on pattern analysis and machine intelligence 39.6 (2016), pp. 1137–1149.
[12] Wheyming Tina Song, Chou Lai, and Yi-Zhu Su. “A Statistical Robust Glaucoma Detection Framework Combining Retinex, CNN, and DOE using Fundus Images”. IEEE Access (2021).
[13] Koen EA Van de Sande et al. “Segmentation as selective search for object recognition”. 2011 International Conference on Computer Vision. IEEE. 2011, pp. 1879–1886.
[14] Yuanbin Wang et al. “A smart surface inspection system using faster R-CNN in cloud-edge computing environment”. Advanced Engineering Informatics 43 (2020), p. 101037.
[15] Beibei Xu et al. “Automated cattle counting using Mask R-CNN in quadcopter vision system”. Computers and Electronics in Agriculture 171 (2020), p. 105300.
[16] 王鈞奕.“利用卷積神經網路對空拍影像中車輛進行偵測與分類”(2018).
[17] 林有德“第三代半導體材料碳化矽發展趨勢研究”.中華科技大學電子工程研究所學位論文(2015), pp. 1–149.
[18] 林泊宏.“基於遮罩區域卷積類神經網路之木節偵測暨分類演算法”.中興大學統計學研究所學位論文(2018), pp. 1–44.
[19] 曾靖堯.“[前瞻性的碳化矽半導體] 長晶製程之自動化與智能化檢測 ”(2020).
[20] 游祿凱 and 黃喬次.“應用人工智慧技術作基板缺陷檢測”.管理學術研討會(2018), pp. 434–443.
[21] 楊雅嵐.“碳化矽元件材料特性比較”.工研院 IEK 電子分項 . 工業技術研究院 , 2010.
[22] Jong Pil Yun et al. “Automated defect inspection system for metal surfaces based on deep learning and data augmentation”. Journal of Manufacturing Systems 55 (2020), pp. 317–324.