Wang, Hai-WeiHai-WeiWangRIH-TENG WU2025-08-282025-08-282025-10-15https://www.scopus.com/record/display.uri?eid=2-s2.0-105012609559&origin=resultslisthttps://scholars.lib.ntu.edu.tw/handle/123456789/731677Building exterior walls often suffer surface damage such as tile spalling due to aging and environmental factors. These spallings may fall and pose severe safety issues to pedestrians and vehicles on sidewalks. However, the current inspection process relies on visual assessments conducted by trained engineers, which is time-consuming and labor-intensive. In recent years, deep learning-based approaches have rapidly emerged and are extensively used for autonomous condition assessments of buildings due to their ability to learn representative features. However training a supervised learning model typically requires a large labeled dataset, which is often unavailable for new tasks. Moreover, tile spalling exhibits significant variations in shape and size, presenting challenges for damage segmentation, particularly with limited training samples. In this study, a novel model called the Multi-Scale Branch Fusion UNet (MBF-UNet) is proposed 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. To minimize the bias caused by the small dataset, repeated trials were conducted to evaluate the performance of the proposed network using a total of 364 manually labeled images from Google Street View or smartphone. Statistical measures have shown that the proposed MBF-UNet surpasses state-of-the-art segmentation models across several general segmentation metrics, while also maintaining low standard deviations. The findings in this study hold significant potential for practical applications in infrastructure monitoring, enabling more effective identification of spalling instances.falseDeep learningNetwork architecturesStructural health monitoringTile spalling segmentation[SDGs]SDG3[SDGs]SDG11journal article10.1016/j.jobe.2025.1135892-s2.0-105012609559