CrackDenseLinkNet: a deep convolutional neural network for semantic segmentation of cracks on concrete surface images

Abstract

Cracks are the defects formed by cyclic loading, fatigue, shrinkage, creep, and so on. In addition, they represent the deterioration of the structures over some time. Therefore, it is essential to detect and classify them according to the condition grade at the early stages to prevent the collapse of structures. Deep learning-based semantic segmentation convolutional neural network (CNN) has millions of learnable parameters. However, depending on the complexity of the CNN, it takes hours to days to train the network fully. In this study, an encoder network DenseNet and modified LinkNet with five upsampling blocks were used as a decoder network. The proposed network is referred to as the “CrackDenseLinkNet” in this work. CrackDenseLinkNet has 19.15 million trainable parameters, although the input image size is 512 × 512 and has a deeper encoder. CrackDenseLinkNet and four other state-of-the-art (SOTA) methods were evaluated on three public and one private datasets. The proposed CNN, CrackDenseLinkNet, outperformed the best SOTA method, CrackSegNet, by 2.2% of F1-score on average across the four datasets. Lastly, a crack profile analysis demonstrated that the CrackDenseLinkNet has lesser variance in relative errors for the crack width, length, and area categories against the ground-truth data. The code and datasets can be downloaded at https://github.com/preethamam/CrackDenseLinkNet-DeepLearning-CrackSegmentation .