Abstract: A road defect detection algorithm SGBNet, which improves YOLOv8, is proposed to address the issues of low detection accuracy, high missed and false detection rates, and poor generalization ability of small targets with road defects in complex backgrounds. Firstly, the Neck part replaced PANet with BiFPN weighted bidirectional feature pyramid to improve the feature fusion ability of the model. Secondly, GAM is introduced into Neck to adjust attention in the feature fusion stage and improve the detection accuracy. Finally, a small target detection layer is added to further enhance the combination of deep semantic information and shallow semantic information. and improve the ability to detect small objects with road defects. Compared with the original YOLOv8n algorithm, the accuracy, recall, and average accuracy of the algorithm SGBNet has been improved by 3. 3%, 2. 5%, and 2. 5%, respectively, achieving more accurate detection of road defects.