To enhance the performance of road defect detection methods, YOLOv8s is used as the benchmark model in this study, with improvements made to the model structure. Firstly, to better detect small object defects, the backbone network of YOLOv8s was modified by replacing the original Conv modules with SPD-Conv, and a new C2f-SPD module was constructed. Additionally, to better utilize the information in feature maps, the original model’s feature fusion module was improved by substituting the nearest neighbor interpolation up-sampling module with a DySample up-sampling module. To better leverage the global context information of images and enhance the model’s generalizability, the LSKA attention mechanism was introduced, creating a new SPPF-LSKA module. To adapt well to complex detection scenarios and environmental backgrounds, DyHead was used to improve the original model’s detection head. The modified model significantly outperformed other models in the task of road surface defect detection. The improved LD-YOLOv8 model achieved a 1.8% increase in accuracy, a 0.9% increase in recall rate, and a 2.4% increase in mAP50.