|
Title |
A Comparative Study on the Performance and Inference Speed of Deep Learning Models for Structure Crack Detection
|
|
DOI |
https://doi.org/10.11112/jksmi.2025.29.6.199 |
|
Keywords |
균열 탐지; 인공지능; 딥러닝; 경량화 모델; 건축물 외관 점검 Crack detection; Artificial intelligence; Deep learning; Lightweight model; Building facade inspection |
|
Abstract |
This study compares and analyzes the performance of deep learning-based image processing models for automatic crack detection on building exteriors. To overcome the limitations of traditional visual inspections, which are time-consuming, labor-intensive, and prone to subjectivity, various CNN- and Transformer-based object detection models were evaluated in terms of detection accuracy, inference speed, and model size. The study particularly focuses on the practicality of lightweight models for real-world applications, experimenting with YOLOv11, Faster R-CNN, EfficientDet, and RT-DETR models. Using a public crack dataset, the results show that the YOLOv11 series achieves a high level of accuracy while maintaining fast inference speeds, making it suitable for real-time field inspections. Additionally, a novel Trade-off Score (TO-score) metric was proposed to quantitatively balance accuracy and speed for practical model selection. This research provides valuable insights for optimizing crack detection models in environments with limited computational resources.
|