Abstract: In response to the high computational complexity, low detection accuracy, and the issues of missed detection and false alarms in the current strip steel surface defect detection algorithm, a lightweight strip steel surface defect detection model, HSED-YOLO, is proposed. Initially, the original YOLOv8n backbone network is replaced with the improved HGNetV2, reducing the redundancy in feature map computation and thereby decreasing the number of the model’s parameters. Subsequently, to further reduce the model’s complexity, a Slim-Neck structured design is introduced into the model’s bottleneck network structure. Concurrently, an EMA attention mechanism is introduced during the feature fusion stage to enhance the model’s feature extraction capability. To further improve the model’s detection accuracy, a DIoU loss function is designed. Extensive experiments are conducted on the strip steel defect dataset. The number of improved model’s parameters and computational load are 2.1×10⁶ and 6.1×10⁹ FLOPs, respectively, which are only 70% and 75.3% of those of the baseline model. Moreover, the average accuracy is improved by 2% compared with the baseline model. These results demonstrate the effectiveness of the improved network.

Key words: defect detection, strip steel, YOLOv8, attention mechanism, loss function, image recognition