基于机器视觉与深度学习的烟用商标纸质量评价方法研究

Research on Quality Evaluation Method of Cigarette Trademark Paper Based on Machine Vision and Deep Learning

关键词:  机器视觉  深度学习  YOLOv8  商标纸质量检测  相似度计算

Key Words:machine vision  deep learning  YOLOv8  cigarette trademark paper quality inspection  similarity computation

摘要:为研究融合深度学习的机器视觉技术应用于烟用商标纸质量检测结果的评价，提出了一种综合评价方法。通过搭建高分辨率工业相机、定制光源及专用软件系统，构建图像标注数据集，采用动态阈值ORB特征检测、改进RANSAC配准优化及多波段融合策略，有效消除拼接缝，图像PSNR达38.9 dB，SSIM为0.94。在要素识别任务中，基于YOLOv8模型引入CBAM注意力模块，结合ResNet-34骨干网络与FPN多尺度特征融合，测试集上mAP50达99.4%，召回率与精确率分别达99.6%和99.0%。设计双分支Siamese网络，融合SIFT描述子与深度语义特征计算相似度，小盒商标纸要素相似度平均准确率97.64%，条盒商标纸相似度平均准确率95.85%。实验表明，该方法显著提升了商标纸质量检测的效率和准确性，为烟草行业提供高效质检解决方案。

Abstract:To investigate the application of machine vision technology integrating deep learning in evaluating the quality inspection results of cigarette trademark paper，this study proposes a comprehensive evaluation method. A high-resolution industrial camera, customized light sources, and specialized software systems were employed to construct an annotated dataset. Dynamic threshold ORB feature detection, optimized RANSAC registration, and multi-band fusion strategies were adopted, effectively eliminating stitching seams. The images PSNR of 38.9 dB and SSIM of 0.94. For feature recognition, the YOLOv8 model was enhanced by introducing a CBAM attention module, combined with a ResNet-34 backbone network and FPN multi-scale feature fusion, achieving 99.4% mAP50, 99.6% recall, and 99.0% precision on the test set. A dual-branch Siamese network was designed to compute similarity by fusing SIFT descriptors and deep semantic features, achieving average recognition accuracies of 97.64% for small box trademark papers and 95.85% for carton trademark papers. Experimental results demonstrate that this method significantly improves the efficiency and accuracy of trademark paper quality detection, providing an efficient quality inspection solution for the tobacco industry.