Road Crack Detection and Classification Using UAV and Deep Transfer Learning Optimization

Unmanned aerial vehicles (UAVs) are effective tools for detecting road cracks. Equipped with advanced sensors and high-resolution cameras, UAVs capture detailed imagery of road surfaces. This data is analyzed using complex image processing and deep learning (DL) techniques to detect and classify anomalies like cracks. Recently, automated methods have improved efficiency and accuracy, replacing traditional visual inspections. DL approaches, trained on large datasets, quickly identify high-performing road cracks. The Deep Learning Assisted Rapid Road Crack Detection and Classification (DL-RRCDC) technique uses ensemble learning and hyperparameter tuning to enhance accuracy. It employs Gaussian filtering to remove noise and the YOLOv8 detector with a squeeze-and-excitation densely connected network (SE-DenseNet). Hyperparameter tuning of SE-DenseNet is optimized using the Northern Goshawk Optimization (NGO) algorithm. Identified road objects are classified using an ensemble classifier with long short-term memory (LSTM), bidirectional gated recurrent unit (Bi-GRU), and autoencoder (AE), with hyperparameters chosen by the dung beetle optimization (DBO) algorithm. Simulations demonstrate the superior performance of the DL-RRCDC technique.