Pavement crack detection and characterization using deep learning and pixel level segmentation

Abstract

Pavement cracks pose a significant threat to road safety and infrastructure integrity, leading to potential hazards for vehicles and necessitating costly repairs. The traditional methods for detecting and characterizing these cracks are often manual, time-consuming, and subject to human error, making it challenging to maintain roads efficiently and effectively. This study advances pavement maintenance by applying deep learning and pixel-level segmentation to improve the detection and characterization (classification and sizing) of pavement cracks, crucial for road infrastructure safety and longevity. The research began with extensive data collection, using a GoPro Hero 8 mounted on a vehicle to gather images of roads in Kuala Lumpur and Selangor, Malaysia. Following detailed preprocessing, including image augmentation, blurring, and light intensity variations, a pavement crack dataset was prepared for analysis. The investigation started with a customized YOLOv7 model, achieving 0.9545 recall and 0.9523 precision on the RDD2022 dataset and 0.9158 recall and 0.93 precision on our custom dataset. When benchmarked against traditional methods, such as ConvNets and deep neural networks, the customized YOLOv7 model demonstrated better precision and recall performance. Subsequent work with the YOLOv8x-seg model resulted in improved precision and recall in crack detection, with performance metrics of 0.93 recall and 0.91 precision for alligator cracks, 0.95 recall and 0.84 precision for longitudinal cracks, and 0.806 recall and 0.89 precision for transverse cracks. In direct benchmarking, this model surpassed previous fusion-based deep convolutional methods in both precision and recall. The final phase involved creating an Advanced Hybrid Deep Learning Model incorporating Deep Gradient ResNet and a Modified Attention mechanism, enhancing crack detection and characterization. This model showed a promising precision of 0.84 for alligator cracks, 0.89 for longitudinal cracks, and 0.87 for transverse cracks, with recall rates of 0.96 for alligator cracks, 0.88 for longitudinal cracks, and 0.80 for transverse cracks. The developed model outperformed the benchmarked research utilizing CrackNet model in both precision and recall. Utilizing advanced segmentation and machine vision techniques, the study successfully demonstrated the capability to precisely size pavement cracks, which is vital for targeted maintenance. The research signifies progress in automated pavement crack analysis, contributing to safer and more durable road infrastructures.