Skip to content

Keras implementation of 'Learning Iterative Processes with Recurrent Neural Networks to Correct Satellite Image Classification Maps'

Notifications You must be signed in to change notification settings

gaudetcj/IterativeSegmentation

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

9 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

IterativeSegmentation

Keras implementation of 'Learning Iterative Processes with Recurrent Neural Networks to Correct Satellite Image Classification Maps'

This paper uses a 2 stage process of training where th results from a CNN are used to train a RNN for refinement. The RNN is analogous to a diffusion process in the form of a PDE. The RNN takes advantage of a MLP being a universal function approximator to solve the system in a x_{t + 1} = x_{t} + \delta x_{t}.

This implementation uses a U-Net like model to generate heatmaps and then a RNN is built by stacking layers that share weights.

#Results Sample results can be seen in the sample-results folder. Note that the image on the left is the segmentation generated by my own CNN from my UltraSoundNerveSegmentation repo. The right is the result of taking that segmentation and the original image and running it through the iterative RNN for refinement.

It is clear that the RNN is reducing noise in the segmentation and even clears up false positives as seen in Figure 8.

About

Keras implementation of 'Learning Iterative Processes with Recurrent Neural Networks to Correct Satellite Image Classification Maps'

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages