人們往往需透過建築的招牌、特殊造型等特徵才能辨別建築。當人們從無法看見建築特徵的立面觀看或建築本身外觀相近時，可能需要花費較長時間才能夠正確辨識建築。本研究旨在訓練一個建築分類模型，並進一步輔助開發應用程式，協助使用者能快速辨識建築物。我們收集國立中央大學校園內的四棟建築外觀影像作為目標資料，以MobileNetV2作為基礎模型，利用遷移學習中的層轉移技術訓練模型，在多元分類的任務上達到95%的準確率。我們比較了不同凍結層數、資料增廣方法與訓練時期數對模型的影響，發現凍結層數的選擇對模型效能影響顯著。另外，我們以人類與Teachable machine進行相同分類任務的結果作為比較基準，人類的準確率為62%，模型之準確率相較之下有明顯提高；Teachable machine的準確率為90%，模型之準確率與其比較則是些微提升。;Identifying buildings for human beings rely on features like signs and distinctive shapes. However, correctly recognizing structures becomes challenging when these features are not visible from a particular facade or when buildings share a similar appearance. This study focuses on developing a deep learning-based model for building recognition that can be integrated into a mobile application, allowing users to identify buildings quickly. The model employs MobileNetV2 as the base model for the layer transfer technique. We collected exterior images of four buildings at the National Central University in Taiwan for training and testing datasets. To optimize the model′s performance, we conducted a parametric study that explored the impact of various factors, including the number of frozen layers, data augmentation techniques, and training epochs. Our model achieved as high as 95% of accuracy for the multi-class classification task. In addition, we conducted the same experiment by humans and Teachable Machine, a web-based machine learning tool developed by Google, as benchmarks for comparison. The accuracy rate of humans was 62%, while Teachable Machine achieved an accuracy rate of 90%. Our model surpassed both of them in terms of accuracy.
