Hybrid connection network for semantic segmentation

In recent years, deep convolutional neural networks like ResNet and DenseNet with short cut connected to each layer can be more accurate and easier to train. Although deep convolutional neural networks show their strength in many computer vision tasks, there is still a challenge to get more precise per-pixel prediction for semantic image segmentation task via deep convolutional neural networks. In this paper, we propose a hybrid connection network architecture for semantic segmentation which consists of an encoder network for extracting different scale feature maps and a decoder network for recovering extracted feature maps with the resolution of the input image. This architecture includes several skip connection paths between encoder and decoder. The paths help to fuse both localization information and global information. We show that our architecture can be quickly trained end-to-end without pre-training on an additional dataset and performs comparable results on semantic segmentation benchmark datasets such as PASCAL VOC 2012.