readme

matlab官方网站中的用于图像融合技术的contourlet变换源代码

Contourlet Hidden Markov Tree (HMT) toolbox (ver. 1.0)

Duncan D.-Y. Po, June 2003

This directory contains the MATLAB files that fit a hidden Markov tree model
to the contourlet coefficients of images. The files require the contourlet
Toolbox by Minh N. Do, which can be downloaded from here:

http://www.ifp.uiuc.edu/~minhdo/software/

In addition, the Image Processing Toolbox is also required.

Assumptions:
1. For all the functions, the image is assumed to be stored as 8-bit unsigned 
integers (uint8).
2. For all the functions, the image is assumed to be square.

There are 4 main functions in this toolbox: train HMT model, denoise, texture 
retrieval, and generate data according to HMT model.

1.) Training
file: pdfbtrainimagethmt (Pyramidal directional filter bank train image tied
                          hidden Markov tree)
This file trains an HMT model to fit an image. For example, if we have an 
image 'lena.gif' in directory '/dir', then

	[model, stateprob] = pdfbtrainimagethmt('/dir/lena', 'gif', '', 0.01)     

trains an HMT model on that image to within an error in model parameter of 0.01
per tree. The resulting model is in the output 'model'. By default, this 
file uses a contourlet transform with '9-7' pyramidal filter and 'pkva' 
directional filter, decomposing the image to 4 scales with 4 directions in the
two coarser scales and 8 directions in the two finer scales. The HMT model 
assumes the contourlet coefficients to be distributed between two states 
and in each state the coefficients have a mean of zero. These attributes can 
all be modified by modifying the first few lines of this file. The output
'stateprob' contains the probabilities for each coefficient to be in each of
the two states. This output is useful for denoising.
If we have an existing model 'mymodel' already in memory and we want to 
start the training using this model, then we can use

	[model, stateprob] = pdfbtrainimagethmt('/dir/lena', 'gif', 'mymodel', 0.01)  

This can be used to speed up the training process, if the existing model is
available. 
For other available pyramidal and directional filter names for the contourlet 
transform, see functions dfilter and pfilters, respectively, in the 
contourlet toolbox.



2.) Denoising
file: pdfbdenoiseimage (Pyramidal Directional Filter Bank Denoise Image)
This file takes in a noisy image and denoises the image based on a given HMT model
and an estimated noise power. For example, if we have an image noisylena.gif in 
path /dir/, and estimate the noise power to be 0.01 (noise power is normalized to
lie between 0 and 1 always), then we can first train an HMT model on the image 
(see (1) above) to obtain its model and stateprob. Then we can denoise the image
as follows   

	cleanimage = pdfbdenoiseimage( model, stateprob, 0.01, '', '/dir/noisylena', 'gif')

The output cleanimage contains the denoised image. By default, this file assumes
the HMT model is trained using the '9-7' pyramidal filter and the 'pkva' 
directional filter. Other filters can be used by changing the first two lines 
of this file. 



3.) Texture Retrieval
file: pdfbclassify_texture (Pyramidal Directional Filter Bank Classify Texture)
This file takes a query texture image and search through a texture image database 
to find the most similar texture images. 
For example, we have a database of 64 texture images in tdb:

	tdb = {'texture1', 'texture2', ..., 'texture64'}

The image files of these textures (texture1.gif, texture2.gif, ..., texture64.gif)
are located in the path /tdir/. Suppose the HMT models of these 64 textures are 
already stored in the path /mdir/. Now we want to search through this database to find
the most similar textures to a texture65.gif. We can do this by

	kld = pdfbclassify_texture('texture65', 'gif', tdb, '/tdir', '/mdir', 0)

Now suppose we don't have the HMT models of the database textures. We can still perform
texture retrieval and the program will train the HMT models for the database textures
and save them. Suppose we want to save the database texture HMT models in the directory
/mdir2/. Then

	kld = pdfbclassify_texture('texture65', 'gif', tdb, '/tdir', '/mdir2', 1)

Note that the output kld is an array of Kublick-Liebler distance between the query
texture and every texture in the database in the order specified in tdb. i



4.) Generate Data
file: pdfbgentdata (Pyramidal Directional Filter Bank Generate Tree Data)
This file uses an existing model to generate random data in tree structure. 
For example we have an HMT model 'hmtmodel' in memory and we want to generate a random
HMT with root level size of 64, then

	tree = pdfbgentdata(hmtmodel, 64)


Notes:
There are also other utility files in this toolbox. Their usage can be found by 
typing 'help <function name>' in matlab prompt. They are:
contourlet.m
contourlet2tree.m
contournc.m
dump_pdfbimagemodel.m
generate_model.m
load_pdfbimagemodel.m
pdfbcalc_imagekld.m
pdfbcreate_equiv_models.m
pdfbflip_model.m
pdfbtestall_imagekld.m
pdfbtrainthmt.m
tree2contourlet.m
type3detransform.m
type3transform.m
type4detransform.m
type4transform.m         

There are also several mex files that might need to be recompiled, depending on 
the system. This can be easily done by typing on MATLAB command window
>> mex <filename>.cc pdfbthmt.cc tree.cc matrix.cc utils.cc

pdfbcalc_KLD
pdfbest_KLD
pdfbgen_tdata
pdfbprotrain_thmt
pdfbtrain_thmt  
dump_pdfbmodel_to_file
load_pdfbmodel_from_file