摘要
目的针对皮革包装产品外表面容易出现划伤、褶皱、污渍等现象,传统图像处理方法依赖手工特征提取,难以适应复杂的纹理变化,其检测精度和鲁棒性均受到限制。为此提出一种基于深度学习的融合注意力机制和加权双向特征网络的皮革缺陷检测方法。方法以YOLOv5s为基准模型,在其主干网络中引入CA、CBAM注意力机制,以增强模型对关键缺陷特征的关注能力。同时,利用加权双向特征金字塔网络重构其颈部结构,提高特征融合效率。最后,将上述改进方法进行联合集成,以提升检测性能。结果在构建的缺陷皮革数据集上进行训练和测试,结果显示,相较于基准模型,所提方法的Precision值提高了2.73%,Recall值提高了2.68%,mAP@0.5值提高了2.67%,mAP@0.5:0.95值提高了2.63%,F1 score值提高了2.7%。结论该方法缓解了传统目标检测模型在皮革包装缺陷检测任务中存在的预测框定位偏差、漏检和误检等问题,为皮革包装缺陷检测提供了可行的改进方案。
The work aims to propose a deep learning-based leather defect detection method that integrates attention mechanisms with a weighted bidirectional feature network so as to solve problems that the surface of leather-packaging products is highly susceptible to defects such as cut,fold,and stains,and that traditional image processing methods rely on handcrafted feature extraction,making them inadequate for handling complex texture variations,thereby limiting both detection accuracy and robustness.Specifically,based on the YOLOv5s architecture,Coordinate Attention(CA)and Convolutional Block Attention Module(CBAM)were incorporated into the backbone to enhance the model's ability to focus on key defect features.Meanwhile,a weighted bidirectional feature pyramid network was utilized to reconstruct the neck structure,thereby improving feature fusion efficiency.These enhancements were jointly integrated to boost detection performance.Experimental results on a custom-built leather defect dataset demonstrated that,compared with the baseline YOLOv5s model,the proposed method achieves improvements of 2.73%in Precision,2.68%in Recall,2.67%in mAP@0.5,2.63%in mAP@0.5:0.95,and 2.7%in F1 score.The proposed approach effectively mitigates issues such as localization deviation,missed detection,and false detection commonly observed in conventional object detection models,providing a practical and improved solution for leather packaging defect inspection.
作者
白先浪
张群利
辛志强
BAI Xianlang;ZHANG Qunli;XIN Zhiqiang(School of Home and Art Design,Northeast Forestry University,Harbin 150040,China;School of Mechanical Engineering,Northeast Forestry University,Harbin 150040,China)
出处
《包装工程》
2025年第17期232-242,共11页
Packaging Engineering
基金
中央高校基本科研业务费专项资金(2572019BL01)。
关键词
包装产业
皮革缺陷
深度学习
注意力机制
加权双向特征
融合集成
packaging industry
leather defects
deep learning
attention mechanism
weighted bidirectional features
fusion integration
作者简介
通信作者:张群利。
