TCN-Net:融合三重注意力机制与特征聚焦扩散的烟支缺陷检测网络  

TCN‑Net:a cigarette defect detection network integrating triplet attention mechanism and feature focus‑diffusion

作　　者：吴庆华[1,2] 张哲铭 赵德华 WU Qinghua;ZHANG Zheming;ZHAO Dehua(School of Mechanical Engineering,Hubei University of Technology,Wuhan 430068,China;Hubei Key Laboratory of Modern Manufacturing Quality Engineering,Wuhan 430068,China;Wuhan Cigarette Factory,China Tobacco Hubei Industrial LLC,Wuhan 430040,China)

机构地区：[1]湖北工业大学机械工程学院,武汉430068 [2]现代制造质量工程湖北省重点实验室,武汉430068 [3]湖北中烟工业有限责任公司武汉卷烟厂,武汉430040

出　　处：《包装与食品机械》2025年第2期30-38,共9页Packaging and Food Machinery

基　　金：国家自然科学基金项目(51275158);湖北省创新群体项目(2022CFA006)。

摘　　要：针对烟支生产过程中微小微弱缺陷(如刺破、黄斑、油渍、隐形夹沫等)检测难度大、对易混淆缺陷的区分能力不足的问题,提出一种改进的缺陷检测网络TCN-Net。融合三重注意力机制,在通道、高度和宽度3个维度进行特征增强,提高微小目标的检测能力,使刺破等缺陷的mAP@0.5提升2.4个百分点;设计一种特征聚焦扩散结构,优化高层语义与低层空间特征的融合,有效提高易混淆缺陷(如油渍、黄斑等)的区分能力,使其mAP@0.5分别提升2.8,1.9个百分点;采用归一化Wasserstein距离损失函数优化目标定位,提升小目标检测精度。试验结果表明,相较于基线模型YOLOv8,TCNNet的mAP@0.5提高5.4个百分点,综合性能优于SSD,YOLOv5和YOLOv7等主流检测算法。研究为烟草工业的缺陷检测提供更精准的解决方案。To address the challenges in detecting minute and subtle defects(e.g.,punctures,yellow stains,oil spots,and hidden tobacco shred inclusions)during cigarette production and the limited ability to distinguish between easily confusable defects,this study proposes an improved defect detection network named TCN-Net.The network integrates a triplet attention mechanism to enhance features across channel,height,and width dimensions,significantly improving the detection capability for tiny targets.This innovation increases the mAP@0.5 for puncture defects by 2.4 percentage points.A feature focus-diffusion structure was designed to optimize the fusion of high-level semantic and low-level spatial features,effectively enhancing the differentiation of easily confusable defects(e.g.,oil spots vs.yellow stains),with respective mAP@0.5 improvements of 2.8 and 1.9 percentage points.The normalized Wasserstein distance(NWD)loss function was employed to refine target localization and boost small object detection accuracy.Experimental results demonstrate that compared to the baseline YOLOv8 model,TCN-Net achieves a 5.4 percentage point improvement in mAP@0.5,outperforming mainstream detection algorithms including SSD,YOLOv5,and YOLOv7 in comprehensive performance.This research provides a more precise solution for tobacco industrial defect detection.

关 键 词：图像处理 烟支缺陷检测 深度学习 注意力机制 NWD损失函数 

分 类 号：TS47[农业科学—烟草工业]