Detection of tiny defects on industrial surfaces is a critical procedure to guarantee product quality and production safety, and relevant research bears significant engineering value. Based on the YOLOv8n model, this paper proposes a lightweight improved YOLOv8 model for industrial tiny defect detection. The improvements are implemented in two dimensions. In terms of training strategy, the regression branches of the detection head are frozen to retain the localization knowledge from pre-trained weights and mitigate overfitting under small-sample conditions. In terms of network architecture, standard convolutions are replaced with Ghost Convolutions, and C2f modules are substituted with C3Ghost modules to substantially reduce the model’s parameter count and computational cost.Experiments conducted on a self-built industrial defect dataset demonstrate that the improved model reduces the number of parameters and computational load by 42.9% and 37.8% respectively, with negligible performance degradation, and its overall performance outperforms the original YOLOv8n. The experimental results verify that the proposed method achieves model lightweighting while preserving detection accuracy, and effectively enhances the generalization capability under small-sample scenarios.