基于改进深度残差网络的快递包装识别融合算法研究  

Research on Fusion Algorithm of Express Packaging Recognition Based on Improved Deep Residual Network

作　　者：邵占英 刘飞 李丽杰 SHAO Zhanying;LIU Fei;LI Lijie(Department of Information Engineering,Hebei Construction Material Vocational and Technical College,Qinhuangdao 066004,China)

机构地区：[1]河北建材职业技术学院信息工程系,河北秦皇岛066004

出　　处：《现代信息科技》2023年第14期61-65,70,共6页Modern Information Technology

基　　金：河北省教育厅项目(SD2022055);校级重点教研课题(PX-2222338)。

摘　　要：在普通的ResNet50的基础上修改激活函数,选取CeLU作为激活函数,增加了3个全连接层FC1、FC2、FC3;并在其中都加入了Droput,最后采用迁移学习的方式去训练模型。在快递包装图像预处理中采用Triplet相似性度量学习方法进行特征提取和用SIFT特征加以完善,以形成新的融合算法,并得出改进的ResNet50网络模型,整体的准确率为98.67%,精确度为97.67%、召回率为98.67%,F1分数为98.33%。通过最后的检测结果,充分地证明了改进的ResNet50融合算法应用于快递包装图像识别的性能优越,也为图像识别技术在智慧物流行业提供技术经验。This paper modifies the activation function on the basis of ordinary ResNet50,selects CeLU as the activation function,and adds three fully connected layers of FC1,FC2,FC3.It adds Droput to them,and finally uses the transfer learning method to train the model.The Triplet similarity metric learning method is used for feature extraction and the SIFT features are used to improve in the preprocessing of express packaging images to form a new fusion algorithm.Finally,the overall accuracy of the improved ResNet50 network model is 98.67%,the precision is 97.67%,the recall rate is 98.67%,and the score of F1 is 98.33%.Through the final test results,it fully proves that the improved ResNet50 fusion algorithm has superior performance in express packaging image recognition,and also provides technical experience for image recognition technology in the intelligent logistics industry.

关 键 词：深度残差网络 图像识别 融合算法 快递包装 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] F259.2[自动化与计算机技术—控制科学与工程]