Semi-supervised semantic segmentation based on improved generative adversarial networks
During the training process,adversarial semi-supervised semantic segmentation networks likely have poor convergence stability and may not model the remote dependence between pixels.In order to solve these problems,a semi-supervised semantic segmentation network is proposed to apply spectral normalization to generative adversarial networks and coordinate attention mechanisms.The spectral normalization is used to make the discriminator of the adversarial network satisfy the Lipsitz continuity,so as to improve the stability of the training process and avoid the problem of gradient disappearance.In addition,the coordinate attention mechanism is integrated into the segmentation network,so as to enable the network to obtain the dependence between distant pixels and enlarge the receptive field.Compared to the benchmark model,when using the 1/50,1/20,and 1/8 labeled datasets in the PASCAL VOC 2012 enhanced dataset,the proposed method im-proves MIoU by 2.2%,1.4%,and 1.8%,respectively.When the 1/8,1/4,and 1/2 labeled datasets in Cityscapes,the proposed method improves MIoU by 1.9%,2.1%,and 1.3%,respectively.The experi-mental results demonstrate that,in comparison with other semi-supervised semantic segmentation networks based on adversarial learning,the proposed algorithm exhibits superior stability and accuracy in semi-supervised semantic segmentation tasks.
国家科技期刊平台
NSTL
万方数据
