Long Short & Attention U-Nets: Deep Learning-Based Models for Building Cracks Identification

Crack detection in building structures, especially concrete structures, is always an important task in monitoring structural health and ensuring structural safety. It is time-consuming and laborious when using a manual crack detection process, and it is also affected by the subjective judgment of inspectors. In the practice of applying artificial intelligence technology, many crack images with poor continuity and low contrast lead to problems of feature extraction and poor training and testing results. The task of crack identification in images is regarded as an image segmentation problem in this paper. Combining with the frequently used Long Short and Attention mechanisms, the original reference U-Net model is improved, and LSU and AU models are designed. The whole process shows that the proposed LSU and AU methods not only eliminate the strict requirements on features compared with traditional identification techniques but also simplify the data preprocessing and model training process compared with current research. Test results on the open crack damage data set show that the overall performance of the subjective evaluation and objective evaluation of the proposed LSU and AU methods is better than that of the original UN and the results of current research.