Traditional quality inspection of zirconium alloy plates mainly relies on manual methods, which are time-consuming, labor-intensive, and prone to low accuracy. To improve the detection accuracy and efficiency of surface defects on zirconium alloy plates, the network structure of the YOLOv8n object detection model was enhanced, and a lightweight multi-scale fusion defect detection model (LEF-YOLOv8) was proposed. To address the issue of small defect features being easily lost, a C2f-edge detail enhancement module (C2f-EDEM) was designed to extract defect edge features while incorporating spatial information, thereby enhancing feature extraction capability. To overcome the lack of explicit feature selection and attention mechanisms in the original model, a global-local interactive attention (GLIA) module was introduced, improving the integration of local and global semantic context. Furthermore, to mitigate the weak information fusion and high computational cost of the original detection head, a lightweight multi-scale fusion detection head (LMSFD) was proposed, which enhances feature representation while reducing model complexity. Experiments on a self-collected zirconium alloy plate dataset and the NEU-DET public dataset demonstrate that LEF-YOLOv8 achieves mAP@0.5 of 86.7% and 75.9%, respectively, improving by 3.2% and 2.4% over the baseline YOLOv8n, without introducing significant computational overhead. These results indicate that the proposed model improves detection accuracy and is suitable for resource-constrained inspection systems, providing an effective solution for automated surface defect detection of zirconium alloy plates.