Image-based spalling detection of concrete structures using deep learning

Currently, concrete structures are maintained based on routine visual inspection to find apparent damage such as crack, delamination, spalling, rebar exposure, and segregation. However, visual inspection requires a lot of labor and time, while the risk to inspectors is high. Thus in this study, an image-based concrete spalling detection technique has been developed using deep learning. First, a training image dataset is built by scraping images from internet, and they are categorized into Spalling, Concrete Joint and Edge, Intact Concrete Surface, and Etc. The number of training images increases by implementing image augmentation techniques. An image classifier to detect concrete spalling is developed by a transfer learning approach that fine-tunes the existing convolutional neural network AlexNet using the augmented training image dataset. With two overlapped windows sliding with stride of a window size, the developed classifier is implemented in each window. A probability map of spalling is then constructed using the average of two score values of the last softmax layer in the classifier. Lastly, pixels with a score larger than 25 % out of 100 % are marked as spalling. The developed approach is validated for 14 spalling images with and without rebar exposure. The existence of spalling is not missed in any cases, and the recall in the pixel level, which shows the detectability of spalling, is found to be over 80% for all test images. The developed approach can be expanded for different types of concrete damage.