Deep convolutional neural network ensemble for pavement crack detection using high elevation UAV images

Abstract

Crack morphology is a major indicator of pavement distress and can indicate the extent of pavement rehabilitation required. Researchers have investigated the detection of cracks using images captured at close proximity. This is often time-consuming, labor-intensive, and inefficient. This research implemented the weighted ensemble technique for detecting pavement cracks on a pixel level using UAV images obtained at high elevations. The images were trained using five deep convolutional neural network architectures: UNet, Vgg-UNet, Resnet-UNet, Inception-UNet, and PaveNet. The pixel-level crack detection results are combined using the ensemble technique to maximize performance. The performance of the ensemble methodology was evaluated and compared with some of the state-of-the-art networks. The predictions obtained were used to estimate the area, length, and mean width of the cracks in the pavement images. It is worth noting that the proposed system can be applied to a specific road segment. A quantitative index is then proposed for quantifying the level of deterioration present in a pavement section.