Graph-FCN for image semantic segmentation

  • 2020-01-02 06:05:29
  • Yi Lu, Yaran Chen, Dongbin Zhao, Jianxin Chen
  • 22


Semantic segmentation with deep learning has achieved great progress inclassifying the pixels in the image. However, the local location information isusually ignored in the high-level feature extraction by the deep learning,which is important for image semantic segmentation. To avoid this problem, wepropose a graph model initialized by a fully convolutional network (FCN) namedGraph-FCN for image semantic segmentation. Firstly, the image grid data isextended to graph structure data by a convolutional network, which transformsthe semantic segmentation problem into a graph node classification problem.Then we apply graph convolutional network to solve this graph nodeclassification problem. As far as we know, it is the first time that we applythe graph convolutional network in image semantic segmentation. Our methodachieves competitive performance in mean intersection over union (mIOU) on theVOC dataset(about 1.34% improvement), compared to the original FCN model.


Quick Read (beta)

loading the full paper ...