原文A 2017 Guide to Semantic Segmentation with Deep Learning

0. Intro

• mainly use natural/real world image datasets
• why medical images are different from natural images
• dataset: VOC2012, MSCOCO
• metric: mean IOU

1. Problem

1.1 before deep

• before deep

  • textonforest
  • random forest based classifier

• prob

  • classifaction fixed input size
  • pooling layer: discard ‘where’ infomation

• patch classication

  • classification networks usually have full connected layers and therefore required fixed size images.

1.2 current

• FCN(prob1)

  • allow segmentation on any size image

• Pooling(prob2)

  1. encoder-decoder arch
  2. dilated conv

• encoder-decoder

  • encoder: reduces the spatial dimension with pooling layer
  • decoder: recover object details and spatial dimension
  • shortcut connections: help decoder recover the object details better

• dilated conv

  • away with pooling layers

1.3 postprocessing

• CRF postprocessing
  
  • similar intensity pixels tend to be labeled as same class
  • boost scores by 1-2%

2. Models