The Improved Fully Convolutional Network applied in Segmentation and Detection for Pavement Crack
Qian Yang, Wenlong Wu, Yuezhi Yang, Wenchen Wu, Zhihao Su
Abstract
Cracks are a typical disease form of airport pavement and highway pavement, mainly caused by heavy traffic load, complex external environment, and performance decay of road infrastructure. Pavement crack recognition technology is transitioning from traditional manual detection to more efficient automatic detection. In this process, semantic segmentation technology is an essential technical means and breakthrough direction to improve recognition accuracy. This paper designs an improved fully convolutional network, which is based on the traditional fully convolutional network model, but the model adopts a feature fusion strategy in downsampling to reduce the loss of details; a multi-scale fusion strategy is used in upsampling to obtain rich information under different receptive fields; Part of the ordinary convolution is replaced by dilated convolution to increase the receptive field to prevent the loss of edge details. We conducted experiments on the CrackLS315 and CrackTree260 datasets. Under the indicators of Precision, it is 8.8% and 9.2% higher than the original fully convolutional network algorithm on the CrackLS315 and CrackTree260 datasets, and it is also better than other classic algorithms.