The work aims to improve the surface defect detection accuracy of new material floor during quality inspection. The defect images collected were expanded by flipping, horizontal migration and vertical migration, and the defect data set of new material floor was constructed. Based on YOLOv5, a prediction head was added to make the algorithm more sensitive to tiny defects. Next, the Swin Transformer module was applied in the feature fusion layer of the network to form the attentional mechanism prediction head and improve the efficiency of network feature extraction. Then, the SE module was added at the end of the backbone to enable the network to extract useful feature information and improve the model accuracy. The experimental results showed that the proposed method could accurately distinguish the quality of floor and identify four kinds of surface defects including white impurity, black spot, edge damage and bubble gum. The mean average precision for each defect type was 82.30%, which was 6.58% higher than that of YOLOv5 Baseline. Compared with other typical target detection algorithms, it could identify floor surface defects more accurately and quickly. The improved YOLOv5 algorithm can classify and locate the surface defects of the floor more accurately, thus greatly improving the efficiency of industrial quality inspection.