Abstract: To meet the real-time demand of manufacturing in the digital workshop,a lightweight semantic segmentation network named Multi Pyramid Pooling Segmentation Network (MPPSNet) that satisfied both accuracy and real-time was proposed,which realized the semantic segmentation for the goal of digital workshop.The improved MobileNetv2 was used as the encoder of semantic segmentation network in MPPSNet,which effectively reduced the amount of network parameters and improved the real-time per-formance of the overall network operation;the multi-pyramid pooling network was used as the decoder of the segmentation network,which could integrate multiple layers of feature information and improve the accuracy of the network.Tests found that the semantic segmentation effect of MPPSNet was better than that of FCN8 and BiSeNet in VOC2012data set.In the self-building object semantic segmentation data set of the digital workshop,the Mean Intersection over Unions (MIoU) of segmenting human,machine tool,and mobile robot of workshop objects reached 71.8% in MPPSNet and the parameter amount of the entire network was 2.55M,which could meet the accurate and real-time requirements of segmenting workshop objects.

Key words: digital workshop, MobileNetv2, real-time, lightweight, semantic segmentation

摘要： 为了满足数字化车间中生产制造实时的需求,提出一种同时满足准确性与实时性的多金字塔池化轻量级语义分割网络(MPPSNet),旨在实现数字化车间目标的语义分割。MPPSNet网络以改进MobileNetv2作为编码器,能有效减少网络的参数量,提高整体网络运行的实时性;以多金字塔池化网络作为解码器,融合多层特征信息,提高网络的准确性。经过测试,MPPSNet在公开数据集VOC2012上的效果优于FCN8与BiSeNet,基于自建的车间目标语义分割数据集,对人、机床、移动机器人3类车间目标检测的平均交并比(MIoU)指标达到71.8%,整个网络的参数量为2.55 M,能够满足数字化车间目标图像实时准确分割的需要。

关键词: 数字化车间, MobileNetv2, 实时, 轻量级, 语义分割