Abstract: Traditional methods for strip steel surface defect detection often suffer from low efficiency and poor accuracy in complex industrial environments. This study therefore proposed a strip surface defect detection method based on YOLOv8n. The present YOLOv8n model is limited in its accuracy of strip surface defect detection for its loss of local information due to global convolution, insufficient diversity in extracted features and easy overfitting during training. To address this, the following improvements are introduced: ① A Clo-conv Net module is first proposed, which uses contextual local enhancement mechanism to strengthen the local features and improve the model's ability to capture local information. ② We put forward the attention mechanism of frequency coordinate FC-CA. It integrates feature frequency with position information to enrich feature diversity by improving the structure of coordinate attention CA. ③ A confidence penalty is introduced, and the output is smoothed so that rapid convergence can be achieved after the model has learned a strong representation. Results show that the model accuracy was significantly improved as it was validated on small and blurred samples. The average precision (0.5) rate improved by 8.9%, 7.2%, and 6.2% respectively compared with the baseline YOLOv8n model while maintaining competitive inference speed.