新型多尺度磁磚剝落分割模型與無監督式對比學習策略之研發

Abstract

磁磚為台灣建物常見之外牆飾材，但其經常會因老化和環境因素而產生表面劣化，造成磁磚碎片剝落等問題。而剝落的碎片對鄰近人行道上的行人和車輛造成損傷的案例屢見不鮮。近年來不少研究使用深度學習及電腦視覺的方式來對外牆進行結構健檢，其中不乏針對磁磚外牆的脫落以及破損的檢測，而現今的主流方法為使用具人工標記的圖片資料集進行監督式訓練，然而監督式方法需要大量破壞的資料及相應的標記數據，使得資料集的構築十分費時費力，尤其是需要像素級標註的語意分割資料集，這也使得相較於傳統的語意分割資料集，該領域的資料較為有限。因此在本研究中，我們設計了一個名為Multi-Scale Branch Fusion UNet (MBF-UNet)的深度學習模型，該模型使用多個不同感受野(receptive field)的分支來增加影像資訊以解決由於資料量過少而引起的過度擬合問題，數據分析顯示此模型架構在我們設計的六個指標中都能有著最好的表現及較低較穩定的標準差。此外我們還提出了一種新式無監督訓練方法，使得我們在訓練過程中不需要任何剝落的標註資料，甚至不需要剝落的圖像，就能使模型辨識出建物圖片上的磁磚剝落，此法基於不確定性評估的概念，以一個不含剝落的已標注房屋資料集對模型進行訓練使其熟悉正常房屋外觀，訓練後模型將對剝落區域產生較高不確定性因為訓練資料不包含剝落，該不確定性即可被評估並輸出異常分數，另外我們還研發了一種剝落生成技術，能夠在正常的房屋影像中添加人工生成剝落以促進模型訓練，數據分析顯示在AUC, AP及FPR95分數中，此方法可優於其他方法18.4%, 46.6%及31.7%。對比於傳統的監督式學習，此無監督式學習法能將資料及建構時間從200小時減少至1小時，換算成工資約台幣36600元。此研究有助於基礎設施監測的應用，並提高磁磚剝落檢測的資料效率。

Exterior wall deterioration, particularly tile spalling, is a common consequence of aging and environmental degradation in urban environments. These structural impairments pose significant threats to public safety, particularly for pedestrians and vehicles on sidewalks. In recent years, deep learning-based approaches have been leveraged in autonomous methods for building condition assessments owing to their capacity to identify structural anomalies. However, training a supervised learning model typically requires a large labeled dataset, which is often unavailable in engineering domain tasks. Therefore, in this study, we design a novel model called the Multi-Scale Branch Fusion UNet (MBF-UNet) for semantic segmentation of tile spalling. The MBF-UNet incorporates additional branches with different receptive fields and self-attention mechanisms to extract meaningful representations of surface damage. Statistical measures have demonstrated that the proposed MBF-UNet outperforms the state-of-the-art segmentation models in 6 general segmentation metrics. Additionally, we propose a new unsupervised learning framework for anomaly detection of tile spalling. There is no need of images that contains spalling or corresponding labels. The framework leverage uncertainty estimation by training a segmentation model with a labeled dataset consisting of known and given classes excluding spalling. After training, the model identifies spalling area as outlier pixels, i.e., the anomaly, due to higher uncertainty score. Besides, we develop a synthetic pattern namely Spalling Craft for outlier exposure to add some anomaly patterns into the inlier building images to enhance model performance. The proposed approach outperforms the state-of-the-art baselines by approximately 18.4%, 46.6% and 31.7% in AUC, AP and FPR95 score, respectively. Compared to supervised learning, our approach reduces the time of dataset construction from 200 hours to merely 1 hour, which is only 0.5% of the original labeling time, and saves approximately 36600 NTD in cost. The outcomes of our study will benefit practical application in infrastructure monitoring, enhancing data efficiency in tile spalling segmentation.