中国烟草学报2025，Vol.31Issue(3)：48-59,12.DOI:10.16472/j.chinatobacco.2023.T0133
基于ROI与CFlow框架的无监督烟叶异物检测研究
Research on unsupervised foreign object detection in tobacco leaves based on ROI and CFlow framework
摘要
Abstract
[Background and Objectives]Foreign objects in tobacco leaves are often unpredictable,and commonly used supervised object recognition algorithms are only suitable for known types of foreign objects that have been trained.The objective of this study is to achieve real-time detection of zero-sample unknown foreign objects in tobacco leaves,effectively improving material purity.[Methods]Unsupervised image anomaly detection(AD)technology was applied,considering the attention mechanism of the human visual system(HVS)and the Conditional Normalizing Flows(CFlow)framework based on deep learning.The detection model was built by combining regions of interest(ROI)with normal tobacco leaf samples to jointly predict the probability of unknown foreign objects in image pixel blocks.[Results](1)The F1 score for unknown foreign object detection was 94.61％,while the object recognition algorithm YOLO v8 had a detection rate of 100％for known foreign objects and 0％for unknown foreign objects.(2)The proposed tobacco leaf foreign object detection deployment scheme based on computer vision had an average prediction time of 0.26 seconds per image,meeting the real-time requirements.[Conclusion]This scheme demonstratesd stronger robustness to unexpected inputs than common object recognition-based foreign object detection algorithms.It can effectively detect foreign objects in tobacco leaves in real time,offering advantages such as low cost and fast deployment,with good potential for widespread application.关键词
计算机视觉/深度学习/无监督/异常检测/ROIKey words
computer vision/deep learning/unsupervised/anomaly detection/ROI引用本文复制引用
吴迪,顾茜,罗旻晖..基于ROI与CFlow框架的无监督烟叶异物检测研究[J].中国烟草学报,2025,31(3):48-59,12.基金项目
福建中烟中细支卷烟产品质量提升关键核心技术攻关重大专项,福建中烟重点科技项目"基于光谱成像技术的在线杂物智能识别与剔除系统研究及应用"(FJZYKJJH2022ZD055) （FJZYKJJH2022ZD055）
