W-Net

From MyInfoRepo
Jump to navigation Jump to search

note: this page uses references in Cite-style, the plugin I have not yet installed

W-Net is an image segmentation network, originating from a paper from Boston University<ref>X. Xia, B. Kulis (2017) W-Net: A Deep Model for Fully Unsupervised Image Segmentation</ref>. This is an informational page about W-Net and reproducing the paper outcome. W-Net is a culmination of many principles from semantic segmentation.

Core Concepts

U-Net

W-Net resembles U-Net in a pretty straightforward way. U-Net is an encoder and decoder of fully convolutional networks, with all layers being connected. W-Net acts like this, but the shape of the U-Net network is repeated. To understand the W-Net, it can't hurt to look at how U-Net is used to do segmentation on images.

First, a contraction is performed to encode the

Fully Convolutional Networks

Normalized Cuts

(https://people.eecs.berkeley.edu/~malik/papers/SM-ncut.pdf)


W-Net

Down-convolution and up-convolution


References

<references />