A novel transfer learning model for the real-time concrete crack detection

Abstract

The crack is an important index to evaluate the damage degree of concrete structure. While, the traditional algorithms of crack detection have complex operations and weak generalization. The performance of crack detection algorithms based on deep learning has been improved, but it also increases the complexity of network structure. Thus, this paper proposes a simplified real-time network for automatic detection of concrete cracks. Taking advantage of a novel transformer-based detector (DETR) architecture integrating receptive fields attention blocks and a feature assignment mechanism, the proposed network can achieve a model with more accurate detection of cracks. Analyzing the association between crack targets and receptive fields on feature layers, the receptive fields attention block is introduced into the backbone, focusing on the target receptive field features. In addition, a neck block is added to the encoder to fuse multi-scale features, with a new feature assignment mechanism assigning features to shallow layers, in order to detect cracks in the shallow features. Finally, given the effect of the inconsistent cracks size on the precision of the loss function, a novel adaptive loss function is applied to replace the loss in the original model. In this paper, the improved model is applied on the homemade crack dataset, and ablation study are done on the added modules to verify their effectiveness. Also, our model is compared with advanced models in cracks detection by using a TIDE toolbox. It is proven that the proposed model has a good effect on crack detection, and has better performance compared with state-of-the-art algorithms.