摘要
电子产品随着更新迭代,对零部件胶体质量的要求不断提高。针对补强胶识别分割,传统算法鲁棒性较低,深度学习的语义分割网络Unet分割速度较慢。为此,提出改进的Unet实时语义分割网络fast-Unet。该网络有3个特征提取分支,输出特征图分别为原图大小、原图的1/4大小和原图的1/16大小,每个分支都共享一部分网络权重。并在第二个特征提取网络中加入了通道分割、注意力模块(CBAM)和金字塔池化模块(PPM)。实验结果表明,fast-Unet相较于Unet网络,在MIoU和MPA上都提升了0.07,FPS提高了43.08,单个样本在线检测耗时仅为25ms,显著提升了补强胶胶体在线检测分割效果。
With the renewal and iteration of electronic products,the requirements for colloid quality of electronic product parts are constantly improving.For reinforcement glue recognition and segmentation,the traditional algorithm has low robustness,and the deep learning semantic segmentation network Unet segmentation speed is slow.Therefore,a real-time semantic segmentation network fast Unet based on improved Unet is proposed.The network has three feature extraction branches.The output feature graphs are the size of the original graph,1/4 of the original graph and 1/16 of the original graph respectively.Each branch shares a part of the network weight.In the second feature extraction network,channel segmentation,convolutional block attention module and pyramid pooling module are added.The experimental results show that compared with Unet network,the proposed fast Unet network is improved by0.07 in both MIoU and MPA,43.08 in FPS,and the online detection time of a single sample is only 25ms,which significantly improves the segmentation effect of reinforcement colloid online detection.
作者
薛金超
赵德安
李长峰
张军
陈辉
XUE Jin-chao;ZHAO De-an;LI Chang-feng;ZHANG Jun;CHEN Hui(School of Electrical and Information Engineering,Jiangsu University,Zhenjiang 212013,China;Changzhou Mingseal Robot Technology Co.LTD,Changzhou 213164,China)
出处
《软件导刊》
2022年第3期213-219,共7页
Software Guide
基金
国家自然科学基金项目(31571571)。
关键词
实时语义分割
Unet
补强胶
注意力模块
金字塔池化模块
real-time semantic segmentation
Unet
reinforcing glue
attention module
pyramid pooling module
