A Deep Learning Network for Classification and Visual Deterioration Detection of Concrete Surfaces

Detecting and addressing defects in concrete surfaces helps ensure the safety of structures and individuals. The timely identification of cracks, spalling, or other defects can prevent accidents or structural failures. Visual inspections are commonly conducted while examining the state of concrete surfaces with flaws. However, this approach is difficult, time-consuming, subjective, and tiresome; additionally, it requires access to many parts of a large project's design. One possible way to improve the accessibility and accuracy of concrete surface data while decreasing human bias is to use deep learning (DL)-based computational algorithms for visualizing deterioration in infrastructure. This paper presents a simple and computationally insensitive deep learning network for classifying whether a concrete surface is defective. Initially, data augmentation is used to handle the different sizes of images in deep learning network training. Then, a pretrained SqueezeNet network is adopted and tested using the CODEBRIM dataset. Finally, a semantic segmentation of the defect using a gradient map was presented for visualizing the defect over the surface. The model achieved 95.60% accuracy on the training dataset and 94.46% accuracy on the test dataset. The proposed network has 12 million learnable parameters, which makes it suitable for implementation on portable devices.