Analysis of Convolutional Neural Network-Based Crack Detection: From Traditional Approaches to Advanced Models

---

Abstract

Our civil infrastructure is incredibly important, but it’s constantly threatened by cracks. If these aren’t dealt with, they can seriously damage structures. Historically, we’ve relied on manual inspections, but these are often slow, subjective, and even risky, making them impractical for today’s vast networks of roads and bridges. This paper explores how automated crack detection, powered by software, is changing the game in structural health monitoring. We start by looking at traditional image processing methods, like using Histogram of Oriented Gradients (HOG) for feature extraction combined with Support Vector Machines (SVM) for classification. Our own experiments with an SVM model showed a foundational accuracy of 82.54%. Then, we shift our focus to the revolutionary impact of Convolutional Neural Networks (CNNs). We present a custom CNN model that achieved an impressive 100% accuracy during both training and validation on its dataset. To put these results into perspective, we compare them with cutting-edge models like YOLOv5 and Mask R-CNN, which demonstrate high precision (94.4% and 85% respectively) and recall (95.6% and 77% respectively) even in complex real-world situations. Through this in-depth analysis, we highlight the exciting progress in automated crack detection, emphasizing how deep learning models are excelling at learning features and delivering robust performance, ultimately helping us ensure accurate, scalable, and reliable assessments of structural health.

---

Keywords

Crack Detection, Convolutional Neural Networks, Image Processing, Machine Learning, Structural Health Monitoring, YOLOv5, Mask R-CNN