| 沈忱,吴黎明,王桂棠,张巧芬.不规则图案透明包装袋缺陷的多尺度智能检测[J].电子测量与仪器学报,2025,39(6):112-120 |
| 不规则图案透明包装袋缺陷的多尺度智能检测 |
| Multi-scale intelligent detection of defects in printed patternson transparent packaging bags |
| |
| DOI： |
| 中文关键词: 深度学习 YOLOv8 损失函数 双向特征金字塔网络 |
| 英文关键词:deep learning YOLOv8 loss function BiFPN |
| 基金项目:佛山市南海区创新创业人才团队(第14批)项目、2023 年佛山市促进高校科技成果服务产业发展扶持项目(2023DZXX02)资助 |
|
|
| Author | Institution |
| Shen Chen | School of Electromechanical Engineering, Guangdong University of Technology, Guangzhou 510006,China |
| Wu Liming | 1.School of Electromechanical Engineering, Guangdong University of Technology, Guangzhou 510006,China;
2.Guangdong Provincial Laboratory of Chemistry and Fine Chemical Engineering Jieyang Center, Jieyang 515200, China |
| Wang Guitang | School of Electromechanical Engineering, Guangdong University of Technology, Guangzhou 510006,China |
| Zhang Qiaofen | School of Electromechanical Engineering, Guangdong University of Technology, Guangzhou 510006,China |
|
| 摘要点击次数: 6 |
| 全文下载次数: 7 |
| 中文摘要: |
| 针对现有不规则图案透明包装袋缺陷检测中低对比度及敏感度不足导致的多尺度异常、图案干扰、漏检等问题，提出一种基于YOLOv8s框架改进的YOLOv8s-CBW检测算法，在YOLOv8s的主干网络C2f模块中嵌入坐标注意力机制（CA），增强模型对低对比度、微小缺陷的空间特征定位与精细化辨识能力；通过双向特征金字塔网络（BiFPN）替换原有的PANet结构，优化多尺度特征融合效率；最后，引入动态聚焦的WIoU v3损失函数提升边界框回归精度，替代传统的CIoU损失函数，提升模型对不规则形态缺陷的边界框回归精度与整体泛化性能。实验表明，相较于基准YOLOv8s型，YOLOv8s-CBW在参数量仅增加0.11×106、浮点数基本不变的情况下，在缺陷检测任务中mAP@0.5达到82.2%，提升了1.3%，mAP@0.5:0.95达到49.3%，提升了7.1%；与YOLOv5s、YOLOv6s等主流模型相比，算法的mAP@0.5分别提高2.3%与10.6%，在保持浮点数基本不变的前提下实现更优检测精度。证明通过轻量化改进的YOLOv8s-CBW在多尺度缺陷的检测中能够保证效率，显著提升稳定性，为包装袋自动化质检提供可靠解决方案。 |
| 英文摘要: |
| To address issues in existing defect detection for irregularly patterned transparent packaging bags—such as multi-scale anomalies, pattern interference, and missed detections, which are caused by low contrast and insufficient sensitivity—an improved YOLOv8s-CBW detection algorithm based on the YOLOv8s framework is proposed. In this algorithm, a coordinate attention (CA) mechanism is embedded into the C2f module of the YOLOv8s backbone network to enhance the model’s spatial feature localization and refined identification capabilities for low-contrast and minute defects. The original PANet structure is replaced with a bidirectional feature pyramid network (BiFPN) to optimize multi-scale feature fusion efficiency. Finally, a dynamic focusing WIoU-v3 loss function is introduced, replacing the traditional CIoU loss function, to improve bounding box regression accuracy for irregularly shaped defects and enhance the model’s overall generalization performance. Experimental results show that, compared to the baseline YOLOv8s model, YOLOv8s-CBW, with only a 0.11×106 increase in parameters and essentially unchanged GFLOPs, achieved an mAP@0.5 of 82.2% (an increase of 1.3%) and an mAP@0.5:0.95 of 49.3% (an increase of 7.1%) in defect detection tasks. Compared to mainstream models such as YOLOv5s and YOLOv6s, our algorithm improved mAP@0.5 by 2.3% and 10.6%, respectively, achieving superior detection accuracy while maintaining essentially the same GFLOPs. This demonstrates that the lightweight improved YOLOv8s-CBW can ensure efficiency and significantly enhance stability in detecting multi-scale defects, providing a reliable solution for automated quality inspection of packaging bags. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
