摘要
针对物流托盘种类多、形体规则度复杂,以及在工业生产环境下托盘存在被遮挡、光照条件变化等因素影响托盘图像分割效果的问题,提出一种新颖的基于激励压缩空洞卷积(SEDC)改进的U-Net网络,通过对特征通道间的相关性进行建模,强化重要特征,提升物流托盘图像分割性能。在SEDC模块中使用1×1×1卷积进行数据降维与升维大幅降低计算量,利用正常卷积与膨胀率为2的空洞卷积探索不同视野下的图像特征,通过SE模块对不同层的重要程度进行自动学习。实验结果表明:相较于现有的一些经典图像分割算法,本文模型在尽可能保证图像分割性能的情况下大幅降低了模型的计算量,并提升了网络的鲁棒性,为物流托盘图像的智能分割提出了一种新的解决思路。
Due to the various types of logistics pallets,complex shape rules,and the problems of pallets being blocked and changing light conditions in industrial production environment,a novel U-Net network based on squeeze excitation dilated convolution(SEDC)was proposed.By modeling the correlation between feature channels,the important features were strengthened and the segmentation performance of logistics pallet images was improved.Specifically,1×1×1 convolution in the SEDC module was used for data dimensionality reduction and dimensionality upgrade,which greatly reduced the amount of calculation,and image features were effectively explored under different fields of view through normal convolution and hole convolution with an expansion rate of 2,while automatically learning the importance of different layers through the SE module.Experimental results showed that compared with some existing classical image segmentation algorithms,the proposed model greatly reduced the computational burden and improved the robustness of the network while ensuring the performance of image segmentation as much as possible,and was expected to provide a new solution for intelligent segmentation of logistics pallet images.
作者
魏占国
宋娅萍
李亚
WEI Zhanguo;SONG Yaping;LI Ya(College of Mechanical&Electrical Engineering,Central South University of Forestry and Technology,Changsha 410004,China)
出处
《包装学报》
2021年第5期35-41,共7页
Packaging Journal
基金
国家林业和草原局科技成果推广计划基金资助项目(2016-51)
湖南省研究生科研创新基金资助项目(QL20210212)。
关键词
物流托盘图像
卷积神经网络
SEDC
logistics pallet image
convolutional neural networks
SEDC
