Performance Comparison of Deep Convolutional Neural Network-based Semantic Segmentation Models for Crack Detection on Asphalt Pavement

Abstract

Surface cracks are the main signs of structural damage in pavement infrastructure networks, especially in transportation systems around the world, where it is essential to perform timely maintenance. Traditional crack detection techniques suffer from low detection accuracy, environmental sensitivity, and assessment inconsistencies. To solve these problems, this research introduces some of the deep convolutional neural networks (DCNNs) for crack segmentation of asphalt pavement. Four DCNN-based semantic segmentation models are selected because they have versatility and can hold spatial information: Mask R-CNN, U-Net, SegNet, and DeepLabV3+. The models help to capture multi-scale features, allowing for the detection of cracks that can vary greatly in size, ranging from fine hairline cracks to larger structural cracks. All the DCNN models are trained on the Crack500 dataset, which features images of cracks of varying intensities, sizes, and orientations, in order to determine the most robust models for various pavement conditions. For the benchmarking process, segmentation performance measures are used, which are Intersection over Union (IoU) and Dice coefficient. Experimental results show that Mask R-CNN achieved the highest segmentation accuracy up to 89.06 and 94.47 for IoU and Dice coefficient, respectively. This model of DCNN has the highest capability among other DCNNs and has contributed to the further development of automatic pavement condition monitoring.