一种基于轻量级孪生神经网络的绿色仓库环境治理图像相似度检测模型  

An Image Similarity Detection Model for Environmental Governance of Green WarehouseBased on Light-Siamese Networks

作　　者：郭彭鹏 金仙力[1] GUO Peng-peng;JIN Xian-li(School of Computer Science,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学计算机学院,江苏南京210003

出　　处：《物流工程与管理》2023年第9期44-49,共6页Logistics Engineering and Management

基　　金：国家自然科学基金(61876091);南京大学计算机软件新技术国家重点实验室开放课题(KFKT2022B01);南京邮电大学校级科研基金(NY221071)。

摘　　要：为监控和改善仓库环境的卫生状况,对仓库管理人员进行有效的评估,从而确保仓库的绿色运营。文中提出了一种基于轻量级孪生神经网络(Light-Siamese Network,简称L-SN)的相似度检测模型,通过输入两张仓库垃圾清理前后的图像,经过设计的轻量级注意力膨胀网络(Light and Attention Dilated Network,简称LADNet)提取输入图像的特征,最后,进行特征对比输出两张图像相似度检测结果来确定垃圾清理前后的图像是否是同一场景,解决仓库管理人员上传图像作假、效率低等问题,并且在自构建的环境监管数据集上与以VGG、Xception、MobileNet等为主干网络的孪生神经网络模型进行实验对比。最终实验结果表明,L-SN模型对图像预测的准确率是95.30%,精确率是95.13%,召回率是95.89%,证明了L-SN网络模型可以在模型参数量与计算量减少的情况下,能够较好的对仓库环境图像进行相似度检测,减轻云端模型算力负荷,实现对仓库管理人员的工作进行有效评估,进而更好的保证和监管仓库环境的卫生状况。In order to monitor and improve the sanitation condition of the warehouse environment,effectively evaluate the warehouse managers,and ensure the green operation of the warehouse,this paper provides a similarity detection model based on Light-Siamese Network(L-SN).By inputting two images of the warehouse before and after garbage cleaning,the designed Light and Attention Dilated Network(LADNet)extracts the features of input images,and finally makes feature comparison and outputs the similarity detection results of two images to determine whether the images before and after garbage cleaning are the same scene.The problems of false image uploading and low efficiency of warehouse managers are solved,and the twin neural network models with VGG,Xception and MobileNet as the main trunk network are compared on the self-constructed environmental regulatory data set.The final experimental results show that the accuracy rate,accuracy rate and recall rate of L-SN model for image prediction are 95.30%,95.13%and 95.89%,which proves that L-SN network model can better detect the similarity of warehouse environment image and reduce the computing load of cloud model under the condition that the number of model parameters and calculation amount are reduced,achieve effective evaluation of warehouse management staff,and then better guarantee and supervise the environment of warehouse.

关 键 词：绿色物流仓库 相似度检测 LADNet网络 轻量级网络 

分 类 号：F251[经济管理—国民经济]