本論文試圖探究如何在資料集高度不平衡且稀少的情況下，利用深度學習之方法，將火災現場所取得之巨觀以及微觀之導線熔痕進行分類，並以Grad-CAM的方法分析深度學習模型所學習之特徵。
本研究所使用的方法，將使用深度學習中遷移學習之概念訓練模型，同時透過資料增強的方法，擴充並平衡資料集之分布，以提高熔痕識別之效能。經過資料增強、資料清理、模型優化與參數調校後，最佳實驗結果得出巨觀通電痕 F1-Score 89.22%、巨觀熱熔痕 F1-Score 80.85%、微觀通電痕 F1-Score 79.46%、微觀熱熔痕 F1-Score 81.90%，並同步建置可用實務上之應用程式原型，達到輔助現場判決與進一步蒐集資料之目的，也期許這樣的成果可提升實務之效率，以提供相關政策制定之參考。
未來希望能以此為基礎，探討更進一步優化導線金相之識別分法，並投入到更多的應用當中，持續改善實務之工作流程。

The objective of this thesis aims to classify the wire melting marks from fire scenes based on deep learning approaches when the data set is imbalanced and only a limited amount of data is available. The correctness of the results is verified through the Grad-CAM method.
This thesis employs the concept of transfer learning to train models, and balance the distribution of the data set through the method of data augmentation, so as to improve the efficiency of melting mark recognition. After data augmentation, data cleaning, model optimization and parameter fine-tuning, the best experimental results in terms of F1 are: 89.22% for macro electricity mark, 80.85% for macro heat-melting mark, 79.46% for micro electricity mark, and 81.90% for micro heat-melting mark. An application prototype has been built to assist on-site recognition and further data collection. It is hoped that the results can enhance the performance and provide references for policies making.
In addition to laying the foundation for further optimizing the wire melting marks identification method, this thesis also improves task efficiency and government's work flow.

[1] 中華民國內政部消防署全球資訊網 火災統計 https://www.nfa.gov.tw/cht/index.php?code=list&ids=220
[2] 中華民國內政部消防署全球資訊網 修正「火災調查鑑定標準作業程序」、「火災原因調查鑑定書製作規定」、「火災原因調查鑑定書分級列管實施規定」之名稱及規定 https://www.nfa.gov.tw/cht/index.php?code=list&flag=detail&ids=23&article_id=343
[3] Johnson, J. M., & Khoshgoftaar, T. M. (2019). Survey on deep learning with class imbalance. Journal of Big Data, 6(1), 27.
[4] Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. (2002). SMOTE: synthetic minority over-sampling technique. Journal of artificial intelligence research, 16, 321-357.
[5] Elkan, C. (2001, August). The foundations of cost-sensitive learning. In International joint conference on artificial intelligence (Vol. 17, No. 1, pp. 973-978). Lawrence Erlbaum Associates Ltd.
[6] Tan, C., Sun, F., Kong, T., Zhang, W., Yang, C., & Liu, C. (2018, October). A survey on deep transfer learning. In International conference on artificial neural networks (pp. 270-279). Springer, Cham.
[7] Goodman, B., & Flaxman, S. (2017). European Union regulations on algorithmic decision-making and a “right to explanation”. AI magazine, 38(3), 50-57.
[8] Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., & Torralba, A. (2016). Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 2921-2929).
[9] Lin, M., Chen, Q., & Yan, S. (2013). Network in network. arXiv preprint arXiv:1312.4400.
[10] Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2017). Grad-CAM: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision (pp. 618-626).
[11] Springenberg, J. T., Dosovitskiy, A., Brox, T., & Riedmiller, M. (2014). Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806.
[12] Chattopadhay, A., Sarkar, A., Howlader, P., & Balasubramanian, V. N. (2018, March). Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV) (pp. 839-847). IEEE.
[13] ImageNet http://www.image-net.org/
[14] Coco DataSet https://cocodataset.org/
[15] Open Image DataSet https://storage.googleapis.com/openimages/web/index.html
[16] Szegedy, C., Ioffe, S., Vanhoucke, V., & Alemi, A. (2017, February). Inception-v4, inception-resnet and the impact of residual connections on learning. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 31, No. 1).