Deep Learning-Based Cigarette Surface Defect Detection Algorithm

Surface defect detection is a crucial step in cigarette manufacturing, directly impacting production efficiency and product integrity. However, current detection methods are limited to traditional machine learning and image processing techniques, which are slow and have low accuracy. To tackle this challenge, we propose an enhanced YOLOv8-based model called PS-YOLOv8 (Precision YOLOv8 for small object detection), where we improve the YOLOv8 backbone by integrating DINOv2 with a ViT-based architecture and replacing the traditional Self-Attention mechanism with Dilated Attention. This modification increases detection efficiency while maintaining high performance. In the feature fusion stage, we apply Patch Merging operations from the Swin Transformer between the attention blocks in the neck structure of YOLOv8, enabling effective multiscale information fusion and enhancing detection robustness. Since the cigarette dataset is private and unpublished, this study uses a self-built dataset, with rigorous preprocessing and data augmentation, along with extensive experiments to ensure the reliability of the results. The experimental results show that the proposed method outperforms YOLOv8. Experimental results show that our model achieves the best performance in terms of Precision, Recall, and mAP50 compared to existing models such as YOLOv8.