Learning Deconvolution Network for Semantic Segmentation

Intro:

Idea:

1. Unpooling:首先subpooling的时候记住max pooling layer选择的最大值的location,然后在Unpooling层还原回原location,其它像素应该是用0代替,得到sparse的map:

Learning Deconvolution Network for Semantic Segmentation_第1张图片

2. Deconvolution:conv_transpose(params可学习):

3. Post-processing: 使用fully-connected CRF给coarse map作refine

你可能感兴趣的:(深度学习,图像分割)