Optimizing Concrete Defect Classification Model With a Novel Comprehensive Dataset

The safety and durability of infrastructure depend greatly on structural health monitoring (SHM). However, traditional SHM methods are labor‐intensive, time‐consuming, and prone to human errors. These issues can be solved with the help of machine learning (ML) and deep learning (DL). This paper presents the creation and application of a comprehensive, generalized dataset that addresses a significant gap in research on structural defect detection and classification. The dataset, developed using an unmanned aerial vehicle (UAV), contains over 7000 labeled images for detection purposes, and more than 50,000 images across five categories, including cracks, pockmarks, spalling, exposed rebar, and rust, for classification. Utilizing this dataset, we trained various models, including CNN‐based, transformer‐based, and hybrid approaches. Our study extensively compares these models in terms of performance and computational efficiency. Additionally, we propose a novel hybrid model, DefectNet, which achieved peak parameter efficiency. This model significantly reduces computational demand while maintaining high accuracy, demonstrating its potential for practical applications in SHM. The proposed network is further validated through real‐world photos, suggesting potential in real‐world monitoring. The results indicate that the proposed methods surpass traditional inspection techniques and offer a scalable solution for SHM.