Concrete Crack Detection Using Embedded Machine Learning

In this work, we present an embedded AI model for real-time concrete crack detection, considering the trade-offs between model complexity and accuracy. For this purpose, we implemented Convolutional Neural Networks (CNNs) on a 32-bit ARM Cortex-M4 CPU and we analyzed inferencing time, peak RAM usage, flash usage, and metrics such as precision, accuracy, recall, and F1 score. The research provides a quantitative comparison between MobileNet versions 1 and 2, and two CNN models with different numbers of layers, aiming to create an efficient and accurate solution for crack detection on surfaces of concrete structural members used in civil infrastructure.