Segmentation and Quantification Framework for Post-Fire Reinforced Concrete Damage Leveraging an Enhanced Yolov8s-Od Network

Abstract

Current deep learning for detecting fire-damage in RC structures lacks damage quantification. This paper introduces YOLOv8s-OD, which identifies concrete cracks and spalling at pixel level. The network integrates Omni-Dimensional Dynamic Convolution (ODConv) module, Dilated-Mobile Inverted Residual Bottleneck Convolution (D-MBConv) module, and Exponential Moving Average (EMA) to enhance detail capture, fuse low and high-level information, and recognize targets at various scales. Comparative ablation studies show that YOLOv8s-OD achieves a Mean Intersection over Union (MIoU) of 93%, with a Frames Per Second (FPS) of 44.19 and a memory size of only 11.02MB. Moreover, comparisons with state-of-the-art networks and tests on public datasets demonstrate the superiority and generalization of YOLOv8s-OD. Additionally, two morphological operation-based methods are proposed to quantify crack and spalling dimensions, validated against actual measurements, offering valuable insights for post-disaster reinforcement and repair in structural engineering.

Keywords: Reinforced concrete components；fire damages；damage segmentation；damage quantification.