Fully CapsNet for Semantic Segmentation
Fully convolutional networks (FCNs) are powerful models for semantic segmentation. But convolutional networks fail to perform well in recognizing and parsing images with spatial variation. In this paper, a novel Capsule network called Fully CapsNet is proposed. We introduce Capsule to FCN and improve Equivariance of the neural network in image segmentation. Compared with traditional FCN based networks, a trained Fully CapsNet shows robustness in recognizing image pixels with more or less spatial variation. Each capsule layer is connected by dynamic routing algorithm. The effectiveness of the proposed model is verified through PASCAL VOC. Results show that Fully CapsNet outperforms the FCN in understanding both original images and rotated images.
Fully convolutional network Semantic segmentation Capsule network PASCAL VOC
Su Li Xiangyu Ren Lu Yang
School of Automation Engineering,University of Electronic Science and Technology of China,Chengdu 611731,China
国际会议
广州
英文
392-403
2018-11-23（万方平台首次上网日期，不代表论文的发表时间）