Dual-instance automatic segmentation and damage quantification of concrete spalling in RC bridges

Reinforced concrete (RC) bridges are crucial components of transportation infrastructure, yet traditional inspection methods, reliant on visual assessments, are inefficient, hazardous, and costly, impeding their frequent application across large networks. In Peru, for example, 71.5% of bridges remain uninspected, which highlights the pressing need for fast, scalable and efficient inspection solutions for bridges networks. To address these limitations, this study introduces an automated Deep Learning (DL) framework for quantifying concrete spalling. The framework leverages a dual-instance segmentation approach that simultaneously localizes damage and a reference scale, enabling direct and accurate damage quantification. To train the different models, an original dataset of images depicting concrete spalling on actual structural members was collected and annotated. Four convolutional neural network (CNN) architectures, U-Net3, DeepLabV3+, Mask R-CNN, and YOLO11-seg, were evaluated for pixel-level image segmentation of spalling. The evaluation identified two models for the final framework: DeepLabV3+ delivered robust segmentation accuracy (mIoU: 0.82), making it highly reliable for detailed offline quantification, whereas U-Net3 (mIoU: 0.83) provided a highly computationally efficient alternative (44.7 FPS) with strict geometric reconstruction capabilities for the spalling, both of them are suitable for practical implementation. These models were then used to extract key damage parameters, such as spalling area and diameter, which were subsequently used to classify damage levels according to current bridge inspection guidelines. The framework is implemented in a user-friendly graphical interface, offering a promising, end-to-end solution for the rapid and accurate condition assessment of RC bridges.