Denoising with Linearizing Belief Nets


BM3D
LBN (K=1)
LBN (K=10)

Wait for the 2 large gifs to load.

The demo compares the large corruption denoising results for the state-of-the-art BM3D model and the LBNs. The LBN generates more realistic images than BM3D while properly modelling the uncertainty associated with denoising. We can also see that increasing the number of Monte Carlo samples K improves the variety of the distribution modeled.

Read the paper