To address the challenges of low accuracy in urban road defect detection caused by varying defect scales and complex environments, this study proposes an improved detection algorithm named YOLOv8-road. The algorithm incorporates a multi-level perception attention (MLPA) mechanism into the backbone network to capture long-range dependencies and extract rich contextual information, enhancing defect feature representation and enabling the model to focus more effectively on defective regions. In the neck structure, a dilated wrapping residual convolution (DWR_Conv) module is integrated into the C2f structure, forming a new C2f_D module that improves multi-scale feature extraction and facilitates the capture of fine-grained defect information while reducing interference from road surface backgrounds. Additionally, the algorithm employs the WIoU loss function to optimize bounding box regression, increasing the model’s adaptability to various defect types and mitigating the negative impact of low-quality samples. Experimental results demonstrate that YOLOv8-road achieves a mean Average Precision at 50% (mAP50) of 98.5%, with a precision of 96.8% and a recall of 96.0%. Compared to the original YOLOv8n model, these metrics represent improvements of 4.2%, 3.6%, and 4.7%, respectively. The proposed YOLOv8-road algorithm exhibits superior performance in real-world urban road defect detection tasks, meeting the practical requirements of road maintenance applications.