Defects in reinforced concrete structures can pose significant threats to the long-term quality and safety of buildings, making it crucial to ensure timely and accurate detection and assessment of these defects. This study proposes an improved version of the YOLOv8 model - RC-YOLOv8 - to address the limitations of existing deep learning models in identifying defects in complex environments, as well as deployment challenges on resource-constrained devices. This lightweight object detection network is specifically designed for detecting concrete cracks and exposed rebar defects with high precision under low computational loads, even in complex backgrounds. RC-YOLOv8 integrates Adaptive Convolution (AKConv), Dual Convolution (DualConv), and the Convolutional Block Attention Module (CBAM), which together enhance feature extraction and fusion, significantly improving detection accuracy and robustness in complex construction environments. To further enhance postdetection defect management, this study also introduces a multilevel risk warning mechanism integrated with RC-YOLOv8, which provides risk scoring and graded warnings based on defect severity, supporting management personnel in quickly responding and taking appropriate maintenance actions. The experimental results show that, compared to YOLOv8n, RC-YOLOv8 reduces the number of parameters by approximately 438,000 and increases detection precision by 9.2 percentage points, recall rate by 2.6 percentage points, and mAP@0.5 and mAP@0.5:0.95 by 7.6 and 6.5 percentage points, respectively. The risk warning mechanism leverages RC-YOLOv8's defect detection results to score risks at multiple levels and trigger warnings based on defect severity, enabling proactive maintenance actions. This approach integrates lightweight network design with a risk assessment framework, offering new perspectives for improving construction quality management.
