- 學位論文
城市街景紀錄與使用深度學習於變化檢測
Urban Street View Recording and Change Detection Using Deep Learning
摘要
現今時代變遷迅速，我們生活的街景也是日新月異，許多過去的回憶漸漸消逝或者被取代，即使有著Google Street View的幫助，因為其記錄間隔時間較長的原因，導致有些街景還未紀錄就已經消失。所以我們希望能有個大家都能輕易使用的方式，來記錄觀察我們生活的城市，並記錄其變化之處。因此本研究將使用Insta360 X3全景運動相機，來進行南投縣埔里鎮的街景拍攝計畫，並且透過其應用程式Insta360 Studio對影像進行預處理，並透過FFmpeg, Equirec2Perspec等工具軟體轉換與擷取，產生最終的影像作為資料集。最後將資料集送入反捲積神經網路[1]，該網路負責偵測兩張不同時間拍攝的街景影像是否有變化，輸入為一組圖像對，輸出則為該圖像對的變化分布，並產生訓練與測試結果。 依據反捲積神經網路架構所搭建並訓練出來的網路，在VL-CMU-CD資料集中測試有高達92.99%的準確率，而在埔里街景的資料集中，因為受到環境干擾的因素更為豐富，所以精確程度有所下滑，但仍有91.21%的正確率。顯示此網路架構雖然較為簡易，但仍有良好的適應性。 根據訓練結果，產生的變化之處，可以幫助我們找回過往歷史與回憶，同時作為產業選擇、未來發展、都市更新的參照與依據，使得我們能有更好資源與經驗去做規劃。
並列摘要
Rapid changes in the modern world are causing many memories to fade away. Even with Google Street View, the long intervals between recordings mean that many scenes are not preserved. Therefore, our first aim is to find a low-cost method to record urban street view regularly. This paper uses the Insta360 [1] X3 panoramic camera to capture street views in Puli Township, Nantou County. The images are preprocessed using Insta360 Studio, and tools like FFmpeg and Equirec2Perspec are used to generate the image dataset. Our second aim to detect street view changes uses deep learning approach. The dataset feed into a deconvolutional neural network for training and testing, with series tests and experiments designed to optimize the results and achieve better change detection. Our contribution includes collecting about 1TB street view data in Puli township, once per month, for two years. A prototype system is established to classifying street view images as containing any major structural change or not. These street view records can help us recover historical memories and the change detection system can automictically provide renewal information in the city.