Deep Learning Pavement Crack Detection based on Atrous Convolution and Deep Supervision

Abstract

As an important part of highway maintenance, pavement crack detection is essential for safe driving. Due to the complex imaging background of pavement cracks, there are problems such as different crack thicknesses, and low contrast between cracks and background. Automatic pavement crack detection systems still need to be more accurate Detection method. In order to improve the performance of road crack detection, this paper proposes a road crack detection network based on atrous convolution and deep supervision (referred to as AD-Net). The network adds a feature extraction module based on atrous convolution between the encoder and the decoder, and deep supervision is introduced in the decoder stage to improve the network’s ability to extract advanced semantic features and strengthen the network’s ability to learn crack edges and suppress noise interference. We conduct experimental verification on the dataset CRACKFOREST and compare it with the classic image segmentation network SegNet, UNet , the experiment proved that AD-Net has a more accurate detection effect in crack detection, and it also obtains better detection results for images with low contrast.