adapting
所属分类:图形图像处理
开发工具:matlab
文件大小:13004KB
下载次数:304
上传日期:2007-11-26 12:19:26
上 传 者:
xiaoxi6615
说明: 给予稀疏编码的原理将10幅自然图像进行系数分解成为图像基。
(Given to the principle of sparse coding will be 10 coefficients of natural images into separate image-based.)
文件列表:
adapting a wavelet pyramid to natural images\IMAGES.mat (20971720, 2007-07-13)
adapting a wavelet pyramid to natural images\matlabPyrTools\binomialFilter.m (309, 1997-04-27)
adapting a wavelet pyramid to natural images\matlabPyrTools\blurDn.m (1215, 1999-03-26)
adapting a wavelet pyramid to natural images\matlabPyrTools\buildGpyr.m (1931, 1998-04-17)
adapting a wavelet pyramid to natural images\matlabPyrTools\buildLpyr.m (2632, 1998-04-17)
adapting a wavelet pyramid to natural images\matlabPyrTools\buildSFpyr.m (3260, 2000-09-05)
adapting a wavelet pyramid to natural images\matlabPyrTools\buildSFpyrLevs.m (1889, 2001-03-24)
adapting a wavelet pyramid to natural images\matlabPyrTools\buildSpyr.m (2017, 1998-04-17)
adapting a wavelet pyramid to natural images\matlabPyrTools\buildSpyrLevs.m (861, 1997-05-02)
adapting a wavelet pyramid to natural images\matlabPyrTools\buildWpyr.m (2605, 1998-04-17)
adapting a wavelet pyramid to natural images\matlabPyrTools\cconv2.m (1325, 1997-04-27)
adapting a wavelet pyramid to natural images\matlabPyrTools\ChangeLog (13784, 2001-03-28)
adapting a wavelet pyramid to natural images\matlabPyrTools\Contents.m (5359, 2001-03-29)
adapting a wavelet pyramid to natural images\matlabPyrTools\corrDn.m (2241, 2001-03-28)
adapting a wavelet pyramid to natural images\matlabPyrTools\corrDn.mex4 (140703, 1998-04-27)
adapting a wavelet pyramid to natural images\matlabPyrTools\corrDn.mexglx (26678, 2000-12-22)
adapting a wavelet pyramid to natural images\matlabPyrTools\corrDn.mexlx (22398, 1998-04-27)
adapting a wavelet pyramid to natural images\matlabPyrTools\corrDn.mexsol (29556, 1998-04-27)
adapting a wavelet pyramid to natural images\matlabPyrTools\einstein.pgm (65596, 1997-04-29)
adapting a wavelet pyramid to natural images\matlabPyrTools\entropy2.m (579, 1997-04-27)
adapting a wavelet pyramid to natural images\matlabPyrTools\factorial.m (120, 1997-05-06)
adapting a wavelet pyramid to natural images\matlabPyrTools\histo.m (1835, 2001-03-28)
adapting a wavelet pyramid to natural images\matlabPyrTools\histo.mex4 (14698, 1998-04-27)
adapting a wavelet pyramid to natural images\matlabPyrTools\histo.mexglx (8934, 2000-12-22)
adapting a wavelet pyramid to natural images\matlabPyrTools\histo.mexlx (6401, 1998-04-27)
adapting a wavelet pyramid to natural images\matlabPyrTools\histo.mexsol (7304, 1998-04-27)
adapting a wavelet pyramid to natural images\matlabPyrTools\histoMatch.m (858, 1998-05-06)
adapting a wavelet pyramid to natural images\matlabPyrTools\ifftshift.m (407, 1997-04-27)
adapting a wavelet pyramid to natural images\matlabPyrTools\imStats.m (1105, 1997-08-22)
adapting a wavelet pyramid to natural images\matlabPyrTools\innerProd.m (404, 2001-03-28)
adapting a wavelet pyramid to natural images\matlabPyrTools\innerProd.mexglx (6751, 2000-12-22)
adapting a wavelet pyramid to natural images\matlabPyrTools\innerProd.mexlx (4479, 1998-05-06)
adapting a wavelet pyramid to natural images\matlabPyrTools\innerProd.mexsol (3948, 1997-08-19)
adapting a wavelet pyramid to natural images\matlabPyrTools\kurt2.m (551, 1997-08-22)
adapting a wavelet pyramid to natural images\matlabPyrTools\lplot.m (1008, 1997-08-30)
adapting a wavelet pyramid to natural images\matlabPyrTools\lpyrHt.m (233, 1997-04-27)
adapting a wavelet pyramid to natural images\matlabPyrTools\maxPyrHt.m (603, 1997-04-27)
adapting a wavelet pyramid to natural images\matlabPyrTools\mean2.m (97, 1997-04-27)
adapting a wavelet pyramid to natural images\matlabPyrTools\MEX\convolve.c (10523, 1997-09-07)
adapting a wavelet pyramid to natural images\matlabPyrTools\MEX\convolve.h (1863, 1997-09-07)
... ...
This directory contains matlab scripts and subroutines for learning a
wavelet basis from natural scenes. These should enable you to
duplicate the results described in our paper, "Learning sparse image
codes using a wavelet pyramid architecture," published in vol. 12 of
the NIPS conference proceedings (available via ftp://redwood.ucdavis.edu/
pub/papers/nips00.ps.gz).
These routines are built upon Eero Simoncelli's matlab pyramid
toolbox, so you will need to have that installed somewhere in your
path in order to run the program. You may obtain this toolbox from
Eero's web page at http://www.cns.nyu.edu/~eero. You may find it
helpful to first go through Eero's tutorial (in the TUTORIALS sub-
directory), as many of the array structures we use are based on his
conventions.
The main script for running the learning algorithm is sparsepyr.m.
You can run it once you start up matlab by simply typing 'sparsepyr'.
You can initialize a wavelet basis in advance if you like, using
randfilt, or if not one will be generated for you. Read the comments
in sparsepyr.m for more information.
The files f54.mat and f55.mat contain 5x5 filters, 3 and 4 bands (not
including scaling function) respectively, that were learned with the
current parameter settings in sparsepyr.m. You can look at these
with the functions showbasis, showbasisfft, and showtiling.
The main parameters you should play with to see how it affects the
result are the learning rate (eta) and the noise variance
(1/lambda_N). Note that changing beta is tantamount to changing
lambda_N, so I usually leave beta fixed at 2.5 and play with lambda_N.
You probably don't want to change the scale parameter sigma unless you
also change VAR_GOAL, and you shouldn't do this unless you work with
data that has a different variance (in which case you should also
change lambda_N appropriately).
If you want to use your own image set, you should first preprocess
them by whitening/lowpass filtering. I did this by multiplying by
the following filter in the frequency domain:
R(f) = f.*exp(-(f/f_0).^4);
where f is the radius in the 2D spatial-frequency plane, and f_0 is
0.8*f_N, where f_N is the Nyquist frequency. Without whitening, the
sparsification step will require many more iterations (because the
curvature will be extremely high along the high spatial-frequency axes
and low along the low spatial-frequency axes), and without the lowpass
filter the bases are bound to come out looking different for
horizontal/vertical vs. diagonal orientations due to the rectangular
sampling lattice.
Questions to:
Bruno Olshausen (baolshausen@ucdavis.edu) or
Phil Sallee (sallee@cs.ucdavis.edu)
近期下载者:
相关文件:
收藏者: