Semantic Labeling in VHR Images via A Self-Cascaded CNN (ScasNet)

by Yongcheng Liu, Bin Fan*, Lingfeng Wang, Jun Bai, Shiming Xiang, Chunhong Pan.

ScasNet

VGG ScasNet

• The encoder is based on VGG-Net variant (Chen et al., 2015), which is to obtain finer feature maps (about 1/8 of input size rather than 1/32).
• On the last layer of encoder, multi-scale contexts are captured by dilated convolutional operations with dilation rates of 24, 18, 12, 6.
• As a trade-off, we only choose three shallow layers for refinement. Moreover, BN layer is not used in VGG ScasNet.

ResNet ScasNet

The configuration of ResNet ScasNet is almost the same as VGG ScasNet, except for four aspects:

• the encoder is based on ResNet variant (Zhao et al., 2016)
• four shallow layers are used for refinement
• seven residual correction schemes are designed for feature fusions
• BN layer is used.

Finetuning

For initializing the encoder part in ScasNet

• The encoder in VGG ScasNet is finetuned with VGG-Net_variant_caffemodel

• The encoder in ResNet ScasNet is finetuned with ResNet_variant_caffemodel

Caffe

• The Caffe we used to train VGG ScasNet is released on DeepLab_v2.

• The Caffe we used to train ResNet ScasNet is released on PSPNet.

Installation

Please follow the instructions of Caffe, DeepLab_v2 and PSPNet.

The code has been tested successfully on Ubuntu 14.04 with CUDA 8.0.

References

1. Chen, L., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A. L., 2015. Semantic image segmentation with deep convolutional nets and fully connected crfs. In: International Conference on Learning Representations.
2. Zhao, H., Shi, J., Qi, X., Wang, X., Jia, J., 2016. Pyramid scene parsing network. arXiv preprint arXiv:1612.01105.

Citation

We would be very glad if ScasNet is helpful for your research, and please consider citing our paper (arXiv):

    @article{liu2018scasnet,   
      author = {Yongcheng Liu and    
                Bin Fan and    
                Lingfeng Wang and   
                Jun Bai and   
                Shiming Xiang and   
                Chunhong Pan},   
      title = {Semantic Labeling in Very High Resolution Images via A Self-Cascaded Convolutional Neural Network},   
      journal = {ISPRS J. Photogram. and Remote Sensing.},   
      volume = {145},  
      pages = {78--95},  
      year = {2018}   
    }   

Contact

We would be very glad if you have some ideas or questions about ScasNet to share with us, please contact yongcheng.liu@nlpr.ia.ac.cn