Deep learning for crack segmentation: Redundant-to-effective feature transformation

Abstract

Deep learning techniques have demonstrated remarkable performance in crack segmentation, particularly in convolutional neural networks like U-Net. However, the problem of feature redundancy in the network limits the network's ability to discover effective features. To address this problem, a feature correlation loss function (FC-Loss) was proposed. FC-Loss directly supervises the average correlation of the features of the first few layers through deep supervision to encourage the model to capture more independent features, and combines the Dropout technology to screen effective features. In order to verify this method, a dataset Crack2181 containing 2181 crack images was constructed. A series of experiments were carried out on it, and the results showed that combining FC-Loss with Dropout technology can effectively alleviate the problem of feature redundancy, improve the accuracy of crack segmentation and the generalization ability of the model.