Automatic concrete infrastructure crack semantic segmentation using deep learning.

To solve the privation of crack information during feature transmission, we propose an automatic concrete infrastructure crack semantic segmentation method using deep learning. Initially, based on the statistical analysis results of crack images, a multi-stage feature extraction network is created with multi-resolution parallel transmission to extract the crack features. Then, the extracted features are aggregated according to the correlation between segmentation classes and pixels to enhance the localization performance of the model for cracks. Moreover, using statistical analysis results as constraints construct the loss function to optimize the model and overcome the data imbalance issue. Experiments are conducted on a self-made manual annotation dataset, which contains 2000 images from the dam, the bridge, and the spillway tunnel, and our method reach 94.51% Precision, 86.39% Recall, 82.26% Intersection-over-Unions, and 90.27% F1_measure on the dataset. The experimental results show that the proposed method is optimal for the semantic segmentation of crack images. • Automatic segmentation of crack images in multiple concrete scenes using the convolutional neural network. • The multi-stage feature extraction network and the class-based feature augmentation network. • Optimization of model training process based on statistical results of crack images. • The model's Precision and Intersection-over-Unions for crack pixels reach to 94.51% and 82.26%. [ABSTRACT FROM AUTHOR]