Enhanced Structural Damage Detection, Segmentation, and Quantification Using Computer Vision and Deep Learning
Dhathri Meda, Mohammed Mustafa Ahmed, Prafulla Kalapatapu, Venkata Dilip Kumar Pasupuleti
Abstract
In infrastructure monitoring, manual inspections and conventional computer vision techniques have long been the standard for detecting structural damage. Nevertheless, these approaches are frequently constrained by their reliance on human knowledge and susceptibility to effectively handle complex or large-scale data sets. Although early machine learning techniques brought automation, they lacked the accuracy and scalability required to manage many kinds and quantities of damage, especially in dynamic contexts. While certain deep learning methods, such as early Mask region-based convolutional neural networks (R-CNN) models and Faster R-CNN, partially solved these issues, they frequently had to choose between computational viability, speed, and accuracy. This study presents a methodology for quantifying structural damage by calculating the percentage area of damage and pixel-based metrics for various types, including cracks, corrosion, and spalling. The damages are then classified as low, medium, high, and critical, depending on their severity. The different types of damages were detected using a state-of-the-art computer vision instance segmentation approach that employed Mask R-CNN and You Only Look Once (YOLO) models like YOLOv5, v7, and v8. A data set of 6,000 images was incorporated for this purpose, where the sizes ranged from 416×416 to 640×640 pixels to achieve optimal balance between speed, accuracy, and resource utilization for instance segmentation models. The data sets were gathered from damaged sites and online data sets and were annotated in polygon annotation format with damages categorized into the different damage categories: cracks, spall, and corrosion. The best model achieved an accuracy of 85%. This demonstrates an effective, convenient, and affordable approach for structural damage assessment. Future work can be extended by integrating with Internet of Things technology like closed-circuit television (CCTV) cameras, street cameras, drones, and dash cameras or adapt it for real-time monitoring applications in diverse environments.