Poisson Noise Reduction example

This sample program is presented at the PDPTA 2013 and PDPTA 2014

"Poisson Observed Image Restoration using a Latent Variational Approximation with Gaussian MRF" H.Shouno and M.Okada

"Accelaration of Poisson Corrupted Image Restoration with Loopy Belief Propagation" H.Shouno

The presentation slides can see at

http://www.slideshare.net/HAL9801/slideshelf

In this sample, prior implementation is just different from the proceeding. The priro shown in the slide is correct implementation.

The implementation language is R with several libraries, "Matrix", "png". "MyPoisson2D.R" and "MyLBP.R" is my local library source. When you start on the R, set working directory to the repositry directory, and you can run with

source( 'Poisson2D13.R' )

if you'd like to apply original EM base program.

source( 'Poisson2DLBP3.R' )

means LBP approximated version, which is faster than the original one.

If you'd like to use with command-line of several shells, please type

$ R --vanilla < Poisson2D13.R

or

$ R --vanilla < Poisson2DLBP3.R

In the running, you can see the trial number, minimum and maximum Poisson parameters (contrast), prior parameters alpha and h, and optmizing values in the cosole.

After running, you can obtain several result files such like "Pois2D13_02_20_01.RData" The numbers in the filename mean "Pois2D13_[min][max][trial].RData". Thus the example means, the first trial of restoration whose original contrast is [02, 20].

If you'd like see the image, please load the data in the R interpreter.

load( 'Pois2D13_02_20_01.RData' )

The image data is assigned as followings

y0: Original Image (with 1D vector format) y1: Poisson noised Image (with 1D vector format) est: Restored Image (with 1D vector format)

The original image size is 64x64 gray scale image. Hence, you can see the image as

image( matrix( y1, 64, 64 ) )

The image should be shown in rotated. If you'd like to correct, please try the rotation the matrix as t(matrix(y1,64,64)[64:1,]).