Open Access

ARTICLE

Intelligent Concrete Defect Identification Using an Attention-Enhanced VGG16-U-Net

Caiping Huang^*, Hui Li, Zihang Yu

School of Civil Engineering, Architecture and Environment, Hubei University of Technology, Wuhan, 430068, China

* Corresponding Author: Caiping Huang. Email:

(This article belongs to the Special Issue: Sensing Data Based Structural Health Monitoring in Engineering)

Structural Durability & Health Monitoring 2025, 19(5), 1287-1304. https://doi.org/10.32604/sdhm.2025.065930

Received 25 March 2025; Accepted 19 May 2025; Issue published 05 September 2025

Abstract

Semantic segmentation of concrete bridge defect images frequently encounters challenges due to insufficient precision and the limited computational capabilities of mobile devices, thereby considerably affecting the reliability of bridge defect monitoring and health assessment. To tackle these issues, a concrete defects dataset (including spalling, crack, and exposed steel rebar) was curated and multiple semantic segmentation models were developed. In these models, a deep convolutional network or a lightweight convolutional network were employed as the backbone feature extraction networks, with different loss functions configured and various attention mechanism modules introduced for conducting multi-angle comparative research. The comparison of results indicates that utilizing VGG16 as the backbone network of U-Net for semantic segmentation of multi-class concrete defects images resulted in the highest recognition accuracy, achieving a Mean Intersection over Union (MIoU) of 80.37% and a Mean Pixel Accuracy (MPA) of 90.03%. The optimal combination of loss functions was found to be Focal loss and Dice loss. The lightweight convolutional network MobileNetV2-DeeplabV3 slightly reduced recognition accuracy but significantly decreased the number of parameters, resulting in a faster detection speed of 71.87 frames/s, making it suitable for real-time defect detection. After integrating the SE (Squeeze-and-Excitation), CBAM (Convolutional Block Attention Module), and Coordinate Attention (CA) modules, both VGG16-U-Net and MobileNetV2-DeeplabV3 achieved improved recognition accuracy. Among them, the CA module (Coordinate Attention) effectively guides the model to accurately identify subtle concrete defects. The improved VGG16-U-Net can identify previously the new untrained concrete defect images in the concrete structural health monitoring (SHM) system, and the recognition accuracy can meet the demand for intelligent defect image recognition for structural health monitoring of concrete structures.

---

Keywords

---