基於BIM模型合成資料與領域自適應提升深度學習模型辨識現地鋼筋影像能力

Abstract

鋼筋混凝土一直是被廣泛使用的建築結構，為保證其結構強度往往需要耗費大量時間與人力進行檢核，近年來有些研究提出使用深度學習與電腦視覺的方法企圖提升鋼筋檢核的效率，然而這些方法大多僅示範在實驗室中搭建的鋼筋籠，亦或是因為不足的鋼筋資料導致模型只能在相似場域上有好表現，本研究提出一套使用合成資料與領域自適應的方法，解決資料不足以及模型表現不理想的情況。 首先透過BIM模型以及自動化腳本，將建構BIM模型與生成合成資料的流程自動化，取得大量與多樣具有實例等級標註的鋼筋合成資料。接著透過Mask R-CNN學習鋼筋的特徵，進行鋼筋影像的辨識，模型學習過程中會使用領域自適應技術消弭合成資料與真實資料間的差距，提升模型的辨識能力。另外因為過去研究並沒有公開的資料集與模型，因此本研究整理與標註現地鋼筋資料，做為本研究數值比較的對象。 本研究最終得到25287張標註的鋼筋合成資料，解決資料不足的問題。在現地鋼筋的測試資料集上，本研究基於合成資料與領域自適應得到的模型比使用現地鋼筋資料做為訓練的模型，在AP50指標上提升超過三倍。此外本研究亦從收集各樣現地鋼筋資料影像供模型預測，本研究提出的模型也都有較好的預測結果，故本研究提出的模型確實有提升現地鋼筋影像辨識能力。

Reinforced concrete has been widely used as a structural material in buildings. To make sure that the structure have enough structural strength, rebar inspection is in need. However traditional rebar inspection takes huge amount of time and involves significant labor. Recently some studies have proposed the use of deep learning and computer vision techniques to enhance the efficiency of rebar inspection. However, most of these methods have only demonstrated their effectiveness on steel cages built in laboratory settings or have been limited by insufficient steel reinforcement data, resulting in models that perform well only in similar scenarios. In this study, we address the challenges of limited data and suboptimal model performance by proposing a method with synthetic data generation and domain adaptation. Large amount and variety of synthetic data has been generated by utilizing BIM and automated scripting. The synthetic data include instance-level annotations. Mask R-CNN model can learn those feature and recognize rebars. We bring domain adaptation into training process to eliminate the domain shift between two different data set, boosting the performance of the model. Since recent related research have not publicize their data set or model, we prepare the on-site rebar data set with annotation ourselves and serve them as a baseline in this research. Finally we developed the rebar synthetic data set with 25287 images, solving the issue of insufficient data. Our model, which have take the advantage of hug synthetic data set and effect of domain adaptation, has achieved triple improvement on AP50 metrics compared to the baseline model when evaluating on on-site testing data set. The model also show greater predictions on the data set we collect from the internet, which is really diverse. Hence, the proposed model indeed enhances the capability of on-site rebar recognition.