本研究為產學合作案，合作廠商是國際第三大矽晶圓材料製造廠，也是台灣碳化矽半導體的龍頭。
碳化矽半導體為矽晶圓製造的新一代材料之一，相較早期的矽半導體材料，有顯著較佳的耐高溫與耐高壓特性，但製程困難與製造成本高。
目前業界檢測碳化矽原料的做法為：先將碳化矽晶圓碇原料切成碳化矽晶圓片，取其中兩片進行破壞性的蝕刻，之後使用特定機台檢測該兩片晶圓切片的缺陷數，以此判斷該晶圓切片所來自的晶碇是否為良品。
本研究動機因為蝕刻為破壞性實驗，經過蝕刻後的碳化矽晶圓片將無法再加工利用，因此提出以非破壞性的檢驗法取代。
為此本研究提出一個修改檢驗流程，其為非破壞性實驗：也就是無須經蝕刻但可以分辨該碳化矽晶圓切片所來自的晶碇是否為良品。
藉由人工智慧中深度學習的 Mask R-CNN模型檢測原本需要由特定機台檢驗的缺陷，以此達到非破壞性的檢驗。
需要檢測的缺陷分為兩類，而本研究針對其中一類缺陷 TSD的檢驗可以達到90%的準確度； TED的檢驗可以達到67%的準確度。
本研究提出非破壞性實驗的修改流程是前瞻性的研究，預計有效大幅節省產學方成本，並提升產學方產品於國內、國際間的競爭力。

This research is University-Industry cooperation with the world's third-largest silicon material manufacturer,
and the leading supplier of silicon carbide(SiC) semiconductor domestically in Taiwan at present.
Silicon carbide semiconductor is the latest silicon wafer manufacturing material, it can support higher pressure and temperature than the earlier silicon semiconductor material.
However, the production process is difficult and the cost of manufacturing is high.
The current practice in the industry for inspecting silicon carbide raw materials is to slice the silicon carbide ingot into wafers first, then etch two of the wafers. After etching,
inspecting the number of defects on the two wafers by a specific machine is the method of determining whether the ingot is a good quality item.
Since etching is destructive, the silicon carbide wafers can not be reprocessed after etching.

This research proposes a modified inspection, a non-destructive inspection.
That is to say, it is possible to determine whether the silicon carbide ingot is a good quality item without etching.
The defects which must have been inspected by a special machine in the past can be detected by using A.I., deep learning model Mask R-CNN in this research.
Thus, it becomes a non-destructive inspection.
Two types of defects must be detected, this research was able to achieve 90% accuracy for one of them, TSD, and achieve 67% accuracy for TED.
This research which is a prospective study provides a non-destructive and modified inspection.
The non-destructive and modified inspection can reduce the cooperation company cost-effectively and improve the cooperation company competitiveness.

致謝...........................................................i
摘要...........................................................ii
英文摘要.......................................................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
1.1.1碳化矽.....................................................1
1.1.2影像辨識...................................................1
1.2研究動機與目的................................................1
1.3符號定義.....................................................2
1.4章節架構.....................................................3
第2章文獻採討....................................................4
2.1 CNN 與 R-CNN................................................4
2.2 Mask R-CNN.................................................5
2.3過往研究的應用...............................................8
第3章 研究方法..................................................9
3.1流程........................................................9
3.1.1 產學方原始流程............................................9
3.1.2本研究提出修改流程用以估計缺陷面積...........................9
3.2蒐集碳化矽晶圓切片資料.......................................10
3.3資料智能化轉换..............................................11
3.4 Mask R-CNN模型訓練.........................................11
3.4.1標記訓練用資料.............................................12
3.4.2更改模型参數..............................................13
3.5計算續效指標................................................14
3.5.1實際績效指標..............................................14
3.5.2估計績效指標..............................................15
第4章研究成果..................................................16
4.1 研究規劃...................................................16
4.1.1電腦設備與環境配置.........................................16
4.1.2 研究資料.................................................16
4.1.3標記測試資料:計算缺陷實際面積...............................17
4.2 續效指標...................................................18
4.2.1 產學方原始績效指標........................................18
4.2.2本研究提出修改後績效指標....................................18
4.2.3 本研究Mask R-CNN模型之評量方式............................19
4.3 Mask R-CNN 檢測成果........................................19
4.3.1 TSD缺陷檢測測成果.........................................19
4.3.2 TED缺陷偵測成果...........................................20
第5章 結論與未來展望............................................22
5.1結論.......................................................22
5.2未來展望...................................................22
第6章參考文獻..................................................24

[1] 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.
[2] 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.
[3] Harnani Hassan, Raudah Abu Bakar, and Ahmad Thaqib Fawwaz Mokhtar. “Face recognition based on auto-switching magnetic door lock system using microcontroller”. 2012 International Conference on System Engineering and Technology (ICSET). IEEE. 2012, pp. 1–6.
[4] Tran Duc Duan et al. “Building an automatic vehicle license plate recognition system”. Proc. Int. Conf. Comput. Sci. RIVF. Vol. 1. Citeseer. 2005, pp. 59–63.
[5] Long Chen et al. “Traffic sign detection and recognition for intelligent vehicle”. 2011 IEEE Intelligent Vehicles Symposium (IV). IEEE. 2011, pp. 908–913.
[6] Yann LeCun et al. “Gradient-based learning applied to document recognition”. Proceedings of the IEEE 86.11 (1998), pp. 2278–2324.
[7] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. “Imagenet classification with deep convolutional neural networks”. Advances in neural information processing systems. 2012, pp. 1097–1105.
[8] 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.
[9] Kaiming He et al. “Mask r-cnn”. Proceedings of the IEEE international conference on computer vision. 2017, pp. 2961–2969.
[10] 謝勻詒. “基於深度學習之頂照式魚眼攝影機下的人物偵測”. 交通大學多媒體工程研究所學位論文 2018 年 (2018), pp. 1–31.
[11] 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).
[12] 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.
[13] Jeremiah W Johnson. “Adapting mask-rcnn for automatic nucleus segmentation”. arXiv preprint arXiv:1805.00500 (2018).
[14] Colin J Burke et al. “Deblending and classifying astronomical sources with Mask R-CNN deep learning”. Monthly Notices of the Royal Astronomical Society 490.3 (2019), pp. 3952–3965.
[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.