橋梁檢測輔助工具精進之研究(1/2)-研訂橋梁檢測3D影像模型作業程序(114銘黃)

本計畫研究對象為梁式橋、板橋、箱型橋等三種類型的混凝土橋梁，研究目標為利用無人機拍攝橋梁構件影像，再透過AI深度學習語意分割技術，偵測影像上橋梁表面各種劣化類型，包括混凝土裂縫、混凝土剝落、鋼筋銹蝕、滲水、白華損傷等。
1. 本計畫以兩種不同形式的橋梁當示範區，包括箱型梁與PCI梁， Y6B1200無人機在自動導航定位的定位誤差，在不使用UWB情況下，絕對定位精確度仍達30~50 公分，可減少人工操作之精神壓力，且可提供高解析影像之品質與穩定度。整體而言，箱型梁橋之高解析影像覆蓋率可達90%，但I型梁橋則受限於主梁間空間太小，高解析影像覆蓋率僅能達到60%。未來若要再提升覆蓋率僅能使用小型無人機，飛入主梁間狹小空間以更近的距離拍照。
2. 本計畫建立一個基於深度學習技術的無人機影像品質評估模型，並製作無人機影像訓練資料集，包括亮度失真、對比度失真、高斯模糊、水平移動模糊和垂直移動模糊等退化影像，所獲得的影像品質評估模型能正確的篩選出品質及格(SSIM≥0.7)和品質不及格(SSIM<0.7)影像。
3. 本年度針對112年所開發之Deeplab V3++橋梁影像劣化辨識模型進行精進，增加了2%高解析影像之標註，用於各劣化模型的微調訓練，使得新模型整體辨識率有提升，F1分數在裂縫、銹蝕、剝落、白華、滲水等劣化類別分別從39.35、25.86、54.33、34.86、11.21提升至53.70、75.41、66.46、69.72、60.85。
4. 在幾何變形偵測部分，本計畫提出兩種方式，針對橋台、橋墩、基礎這類的構件，可以使用不同時期的3D點雲，透過點雲距離的計算分析變形量。而若是主梁或帽梁之類的構件，若有垂直面向下滑動的變形，則可以製作不同時期的正射影像，並透過特徵點匹配技術計算位移量或撓度，用以分析其變形量與變形型態。

ABSTRACT:
The project of this study are three types of concrete bridges: beam bridges, slab bridges, and box girder bridges. The research aims to use UAV to capture images of bridge components. AI deep learning-based semantic segmentation techniques are then applied to detect various types of surface deterioration on the bridge, such as concrete cracks, spalling, rebar corrosion, infiltration, and efflorescence.
1. This project uses two different types of bridges as demonstration areas, including a box girder and PCI girder. The positioning error of the Y6B1200 UAV in automatic navigation, without using UWB, still achieves an absolute positioning accuracy of 30–50 cm. This reduces the mental stress of manual operation while providing high-resolution image quality and stability.
2. This project established a UAV image quality assessment model based on deep learning techniques and created a training dataset of UAV images. The dataset includes degraded images such as brightness, contrast, Gaussian blur, horizontal motion blur, and vertical motion blur. The resulting image quality assessment model can accurately classify images as either passing (SSIM ≥ 0.7) or failing (SSIM < 0.7) in terms of quality.
3. This year, improvements were made to the Deeplab V3++ bridge image defect recognition model developed in 2023. An additional 2% of high-resolution images were annotated and used for fine-tuning the AI model, resulting in an overall improvement in recognition accuracy. The F1 scores for different degradation categories improved as follows: cracks (39.35 → 53.70), rust (25.86 → 75.41), spalling (54.33 → 66.46), efflorescence (34.86 → 69.72) , and infiltration (11.21 → 60.85).
4. Regarding geometric deformation detection, this study proposes two methods. For components such as abutments, piers, and foundations, 3D point clouds from different periods can be used to calculate and analyze deformation amounts by measuring point cloud distances. For components like main girders or cap beams, if vertical sliding deformations occur, orthoimages from different periods can be generated, and feature point matching techniques can be applied to calculate displacement or deflection, allowing for an analysis of deformation magnitude and patterns.

作者簡介：

交通部運輸研究所：
饒見有, 林昭宏, 劉光晏, 李志清, 賴威伸, 胡智超