摘要
目的:为解决软包装产品在出厂检测时速度慢、效率低、误检率高、标准不统一等问题。方法:通过数据预处理和迁移学习,使用ResNet50网络模型框架,对软包装杯沿缺陷检测进行了实验和分析。首先对小样本数据集进行预处理,其次采用迁移学习的方法,将特征提取能力较强的参数引入本地模型,再针对杯沿数据集重新训练提高模型分类精度。结果:在小规模软包装杯沿图像集上神经网络模型准确率可达97.69%。结论:由此可见,该实验设计对解决食品软包装杯沿缺陷分类问题有效。
Aims:This paper aims to solve the problems of slow speed,low efficiency,high false detection rate and non-uniform standard in the ex-factory testing of soft packaging products.Methods:The ResNet50 network models were established after preprocessing the data of the edge defect of soft packaging cups.The models were re-trained after introducing the parameters with strong feature extraction ability into the local model by using the transfer learning to improve the classification accuracy of the model.Results:The results showed that the accuracy of the neural network model was 97.69%to small-scale soft-packaging cup edge image sets.Conclusions:The established model is effective in the detection of edge defects of flexible food packaging cups.
作者
金宇霏
陆慧娟
郭鑫璐
张俊
朱文杰
JIN Yufei;LU Huijuan;GUO Xinlu;ZHANG Jun;ZHU Wenjie(Key Laboratory of Electromagnetic Wave Information Technology and Metrology of Zhejiang Province,College ofInformation Engineering,China Jiliang University,Hangzhou 310018,China;Institute of Food Sciences,Zhejiang Academy of Agricultural Sciences,Hangzhou 310021,China)
出处
《中国计量大学学报》
2021年第3期325-331,共7页
Journal of China University of Metrology
基金
国家自然科学基金项目(No.61272315)
浙江省自然科学基金项目(No.LY21F020028)
现代农业产业技术体系建设专项项目(No.CARS-26-04BY)
浙江省大学生科研创新活动计划项目(No.2021R409054)。
