Lightweight frameworks for real-time crack monitoring in civil infrastructure

Cracks in concrete buildings, pavements and bridges are important signs of structural deterioration and present serious concerns to integrity and safety. Tradionally, manual inspection has been in use to detect the cracks which is labour-intensive, subjective and prone to errors. Recently, researchers have evolved an advanced automated techniques such as YOLO (You Only Look Once), to overcome the manual errors. Continuous refinements have led to the developments of sophisticated versions YOLOv4 to YOLOv11 in the YOLO series. This study suggests a customisation of YOLOv11, for the purpose of its quickness, high accuracy and recall. Using data augmentation, hyperparameter optimisation and transfer learning on a composite dataset of concrete crack images, this model is specifically customized for crack detection. Based on experimental and publicly accessible data like SDNET2018 (Structural Defects Network), this customized version outperforms baseline versions YOLOv5, YOLOv8, YOLOv9, YOLOv10 and YOLOv11. An mAP50 (mean Average Precision) value of 68.6% is achieved which is 3.47% higher as compared to YOLOv11. Similarly, a precision of 80.8% and recall of 63.6% are achieved. The study provides 50, 100, 200, 300 and 400 epochs for training and validation. The 100 layers and 6.3 GFLOPs (Giga Floating Point Operations Per Second) of this model are also very less compared to other given models which is an indicator of less complex model. This model has proved computationally efficient and suitable for real-time applications and robust to challenging conditions such as low contrast and complex backgrounds, making it a valuable tool for structural health monitoring.