Abstract: Objective To improve efficiency and accuracy of weld surface quality inspection, a steel structure weld surface defect detection system based on YOLOv8 deep learning model is proposed in the paper. Methods By improving network structure and training strategy, the model optimizes detection performance in industrial scenarios. It adopts CSPDarknet53 backbone network integrated with C2f module to enhance multi-scale feature fusion capability. It introduces an anchor-free decoupled detection head structure to decouple classification and regression tasks, reducing the model’s dependence on prior parameters. It combines Varifocal Loss (VFL) and Distribution Focal Loss (DFL) to optimize loss function, improving detection sensitivity for small target defects. Results In the experiment, ZEDmini camera was used to collect real samples, and a self-built steel structure weld dataset was created. Among them, recognition accuracy of good welds reached 96.2%, and detection accuracy of defective welds stood at 82.5%. The training process curves show that box_loss and cls_loss converged after 120 and 150 epochs respectively. The mAP50 index of validation set stabilized at 76.7%, and mAP50-95 index stabilized at 65.5%. Test results indicate that the model achieves an inference speed of 45 frames per second in 640×640 resolution images, inference time per image is approximately 22.2 ms. Conclusion The steel structure weld surface defect detection system based on YOLOv8 deep learning model can meet requirements of real-time detection and provide reliable technical support for intelligent inspection of welding quality.