Efficient Crack Detection in Concrete Structures Using Dual-Channel Deep Convolutional Networks

Due to repetitive loading and the ensuing fatigue stress, cracks develop in civil structures, compromising their safety. Modern advancements now incorporate machine vision to enhance the maintenance, monitoring, and inspection of concrete infrastructures, thereby reducing the reliance on manual, on-site inspections. This paper introduces a crack detection methodology utilizing a deep convolutional neural network (CNN) that identifies concrete cracks without the need to explicitly compute defect features. Throughout this research, a dataset of 3200 labeled images featuring various concrete cracks was developed, capturing a wide range of contrasts, lighting conditions, orientations, and crack severities. Initially employing a deep CNN trained on these 256 × 256 pixel-resolution images, the model was progressively refined by addressing identified challenges. Enhancements included an augmented dataset simulating conditions typical of drone-captured footage, such as random zooming, rotation, and intensity scaling, alongside comprehensive ablation studies. This culminated in the development of a dual-channel deep CNN, which demonstrated high accuracy (approximately 92.25%) and robustness in detecting concrete cracks under realistic conditions. The model's performance was rigorously tested and further validated through analysis of feature maps, underscoring the effectiveness of the dual-channel architecture.