Architecutures for Image segmentation

Fully Convolutional Network (FCN)

  • This is a typical encoder-decoder architecture.
  • Encoder compresses the information
  • Decoder will upsample and use skip connections to come up with orginial image
  • This image is used for segmentation

FCN Architecture FCN research paper

UNet

UNet Architecture Unet research paper

PSPNet (Pyramid Scene Parsing Network)

PSPNet Feature Pyramid Network Research paper

Deeplab

Deeplab

Deeplabv3+

Deeplabv3+ Atorous convolution with Dilation rate Research paper Conditional Random Fields