面向无监督分割的双分支上采样域自适应网络

Double branch upsampling domain adaptive network for unsupervised segmentation

工业应用中,表面压印字符图像全监督语义分割将会给企业带来高昂的数据集标注成本,针对该问题,提出了双支路特征融合的域适应分割方法(dual-branch feature fusion domain adaptation,DbFFDA)。借鉴U-Net的跨层连接设计思路,提出了双分支上采样结构的残差域适应分割网络(residual adaptation network,Res-Adp);同时提出了融合特征输入用于提升网络分割性能,克服了字符缺失的问题;此外,提出了分割连续性损失函数L_(Con),抑制了分割图像中噪点的产生。在石墨电极表面压印字符无监督分割实验中,所提方法MIoU值可达69.60%,实际分割效果已基本满足字符识别需求,有望在特定工业场景中投入实际应用,为企业节省数据集标注成本。

In industrial applications,fully supervised semantic segmentation of surface-imprinted character images will bring high dataset annotation costs to enterprises.To solve this problem,this paper proposed a dual-branch feature fusion domain adaptation(DbFFDA)method.Firstly,this method drew on the design idea of U-Net’s cross-layer connection,and proposed a residual adaptation network(Res-Adp)with a dual-branch upsampling structure.At the same time,this method proposed the fusion feature input to improve network segmentation performance and overcome the problem of missing characters.Furthermore,this method proposed a segmentation continuity loss function L_(Con),which suppressed the generation of noise in segmented images.In the unsupervised segmentation experiment of imprinted characters on the surface of graphite electrodes,the MIoU value of the proposed method can reach 69.60%.The actual segmentation effect has basically met the needs of character recognition,and it is expected to be put into practical applications in specific industrial scenarios,saving enterprises dataset labeling costs.