A Comprehensive Data Augmentation Method for Modern Cigarette Packaging Defect Detection

This study addresses the issue of insufficient quantity and incomplete types of defect detection datasets for cigarette packages in modern industrial production line scenarios. We propose a cigarette packaging defect data augmentation method. The method consists of three components: preprocessing for obtaining cigarette package defect data through segmentation and cropping based on Segment Anything Model(SAM), generation of cigarette package defects using Cigarette Box Defect-Improved Denoising Diffusion Probabilistic Model (CBD-IDDPM), and We expanded the original dataset by utilizing 600 generated images and trained different defect detection models using the augmented dataset. The experimental results validate the effectiveness of our dataset augmentation method for cigarette package defect detection. The mAP50-95 of YOLOv8n has increased from 73% to 75.2%.