Multi-scale classification network for road crack detection

Abstract

Feature maps of different scales in convolutional neural networks (CNNs) can be regarded as image pyramids. In classification tasks, only the last layer of feature maps is used for making decision. However, in tasks such as road crack detection, the target objects are so small that some original information might lose during the downsampling process in CNN. The authors propose a structure that uses the information that is contained in different layers of feature maps, so all the information could contribute to the classification. This process is managed by adding the weighted values of pixels in corresponding regions of different layers in feature maps and using the sum of these values as the output. The authors apply this structure on a residual network and use it to learn the features of road cracks. Experiments have shown that with the authors' structure, the network performs better than others at understanding and detecting road cracks.