以深度學習與平面投影轉換進行快速鋼筋查驗

Abstract

本研究旨在整合深度學習與基於平面投影轉換的影像拼接技術，以建立一個快速的鋼筋查驗流程。因單張二維影像僅能完成局部查驗，本研究提出透過影像拼接生成大型全域圖像，並在全域圖像上進行鋼筋辨識，以實現全域定位的查驗效益。本研究結合三維重建與二維實例分割，使用稀疏點雲萃取相機資訊，並以鋼筋構件平台作為正射影像平面，以穩固完成二維圖像的拼接。本研究對二維深度學習模型的偵測結果進行後處理，並自動產生結果圖和文字報告，與拼接圖像結合，使工程師能夠一目了然地了解查驗項目的缺失。深度學習模型訓練過程亦引入主動學習，以提升標註流程的效率，降低人力成本。 本研究通過實驗室場域和實務場域的測試，從查驗效率和查驗準確度兩個面向評估此查驗流程的可行性和穩健性。在準確度方面，經現場場域的測試，本研究在雙向版結構與樓梯間牆結構皆取得良好表現，鋼筋支數查全率最高達100%召回率，鋼筋間距誤差最低達到11.05mm、6.7%。此外，通過實驗室場域測試，本研究探討工程師拍攝工法對查驗模型表現的影響，得證對鋼筋垂直拍攝可將模型表現最大化。在查驗效率方面，本研究在實驗室場域最快可於五分鐘內完成查驗，現場場域則為約十分鐘左右，平均每張照片運行時間最快達7.77秒，相較三維重建的鋼筋查驗方法，能節省高達88%的運行時間。

This study aims to integrate deep learning and image stitching techniques based on homography transformation to establish a fast rebar inspection process. Given that a single 2D image is only suitable for localized inspection, this research proposes generating a large image through image stitching and conducting rebar recognition on it to achieve the benefits of large-scale inspection. This study combines 3D reconstruction and 2D instance segmentation, utilizing sparse point clouds to extract camera information and rebar component platform as the orthographic image plane to ensure stable 2D image stitching. Post-processing is applied to the detection results of the 2D deep learning model, and result drawings and text reports are generated, combined with stitched images, enabling engineers to readily understand the inspection results and their defects. Active learning is introduced during the training process of the deep learning model to enhance annotation efficiency and reduce labor costs. The inspection process is tested in both laboratory and practical field scenarios to investigate the feasibility and robustness of the process from the perspectives of inspection efficiency and accuracy. Regarding accuracy, based on tests in practical field scenarios, the process achieves good performance in both bi-directional slab structures and staircase wall structures, with a maximum of 100% Recall and a minimum of 11.05mm and 6.7% error in rebar spacing. Furthermore, through laboratory field testing, this study investigates the influence of shooting methods on the process, proving that perpendicular shooting of rebars maximizes the performance of the model. In terms of inspection efficiency, the inspection can be completed within five minutes in laboratory scenarios and about ten minutes in real scenarios. The average runtime for each photo can be as fast as 7.77 seconds, resulting in up to 88% time savings compared to the 3D reconstruction of the rebar inspection method.