Image Segmentation of Concrete Cracks Using SegNet

Abstract

Inspecting flaws in a structure are vital for engineering applications, especially in concrete projects. The goal of this paper was to employ semantic segmentation model named as SegNet to identify concrete cracks for the continuously and automatically structural health monitoring. The commonly used Adaptive Moment Estimation algorithm and Stochastic Gradient Descent algorithm were applied for optimization. Various recently objective loss functions were served as the evaluation function for image segmentation. Different raw input images of concrete cracks under various conditions such as the shape of cracks, width of cracks, rough or smooth surfaces of backgrounds, were divided for training and validation subsets. The findings revealed that both optimizers performed the similar accuracy by using the intersection over union for concrete crack inspections. In addition, dice, tversky, and focal tversky losses showed better than binary cross-entropy and lovasz losses in terms of the overall accuracy of image classification problems.