Perspective Correction and Deep Learning‐Based Crack Detection for Concrete Structures

Abstract

ABSTRACT This study introduces a practical and easy‐to‐implement method for detecting cracks in concrete structures that meets engineering application requirements. The approach integrates a distance sensor with an image perspective distortion correction algorithm and employs deep learning techniques to automatically correct image distortions, extract crack information, and quantify the crack condition. This method addresses the inadequacies of previous studies in considering perspective distortion, enabling automated and efficient image capture and accurate crack identification. First, we fine‐tuned a lightweight object detection model based on the pre‐trained YOLOv8x model within the YOLOv8 framework, using a custom dataset to enhance its performance. Next, we trained a semantic segmentation deep learning model using several public datasets containing 9584 crack images and their corresponding pixel‐level annotations for precise crack detection. Additionally, a distance sensor combined with a calibration and image processing algorithm was used to remove perspective distortion and convert the size of structural defects from pixels to millimeters. This process includes capturing component edges, calculating the aspect ratio, and performing perspective correction, ensuring high accuracy in images of cracked structures taken from various angles. Field tests showed that this method improves crack detection accuracy and measurement precision under correct conditions. It simplifies the detection process and offers a reliable automated solution, enhancing the efficiency and accuracy of crack monitoring in concrete structures.