Detection of Varieties of Cracks in Concrete Structures Using Selected Machine-Learning Techniques

This paper presents a multimedia image analysis framework for detecting cracks in concrete structures using deep learning models.The motivation for this study stems from the increasing need for automated structural health monitoring to reduce manual inspection costs and enhance safety.Crack images are treated as structural multimedia data, positioning the study within computer vision and content-based image processing.This framing allows for leveraging advanced multimedia processing techniques to extract meaningful features from complex structural imagery.Two architectures, ResNet50 and Vision Transformer (ViT), were applied to a curated dataset of 10,000 crack images categorized into hairline and structural cracks.The dataset was meticulously annotated and balanced to ensure model reliability across different crack types and sizes.Preprocessing and augmentation were implemented as multimedia content transformation steps to enhance robustness under varying imaging conditions.Augmentation techniques included rotation, scaling, and contrast adjustments to simulate real-world inspection scenarios.Experimental results show strong performance, with ResNet50 achieving 99.80% testing accuracy and ViT 99.40%.These results indicate that both CNN-based and transformer-based approaches can effectively capture intricate patterns in structural defects.Beyond structural health monitoring, the approach contributes to multimedia analytics by demonstrating how transformer-based and residual networks can generalize to domains including remote sensing, industrial defect recognition, and automated surveillance.Moreover, the study provides insights into model interpretability, showing which image regions most strongly influence predictions, thus aiding trust in automated systems.The findings support integration into UAV inspections and robotic monitoring systems, advancing intelligent multimediabased structural health assessment.