應用UAV 影像及深度學習技術輔助橋梁裂縫量化分析

Application of UAV Image and Deep Learning to Assist Bridge Cracks Quantitative Analysis

臺灣的橋梁約有兩萬九千座，依據我國公路橋梁檢測之規範，橋梁於完工後每兩年需要進行一次定期檢測，橋梁定期檢測的方式以目視檢查法進行，檢測重點以裂縫為主，但許多裂縫構件位於高空或河面上，故該法需專業橋檢人員藉由操作橋檢工程車等方式，以接近橋梁構件進行檢測，並由人員自行判斷並針對構件劣化的情形予以評分，以最終評估分數判斷是否需要緊急修復。上述之傳統檢查方式之過程不僅成本高、風險高、耗時耗力且受人為主觀限制，容易造成檢測成果不準確。因此本研究擬採用深度學習物件偵測網路YOLOv4模型訓練出一套橋梁裂縫辨識模型，並以UAV拍攝橋梁裂縫影像，後將影像進行裁切再以模型進行逐步辨識，最後採用影像處理邊緣檢測技術Canny及形態學對裂縫影像進行輪廓萃取，再於影像上進行裂縫寬度量測。研究的最終裂縫量測精度優於0.22mm，研究顯示可改善傳統方法的限制，並提升檢測效率。

There are about 29,000 bridges in Taiwan. According to our country's highway bridge inspection standards, the completed bridges need to be inspected regularly every two years. The method of inspecting bridges is usually carried out by visual inspection, and the inspection is mainly based on cracks. However, many cracks are located at high altitude or on the river surface, so this method requires professional bridge inspectors to take bridge inspection engineering vehicles, equipped with slings and take small boats to approach the bridge components for inspection, and the personnel will use In the way of subjective judgment, the situation of component deterioration is scored, and the final evaluation score is used to judge whether urgent repair is required. The above-mentioned traditional inspection methods are not only low-cost, high-risk, time-consuming and labor-intensive, but also because many bridges need to be inspected every year. Therefore, if the traditional inspection method is adopted, the bridges inspected will be delayed and the safety of passersby will be endangered. Therefore, this study intends to use deep learning to establish a set of crack identification models, and use UAV to detect bridge crack in image, then cut the images and identify them step by step with the model. The final fracture measurement accuracy of the study is better than 0.22mm, and the study shows that it can improve the limitations of traditional methods and improve the detection efficiency.