Chi squared two sample test to measure image correspondence
in [Matlab]
|
From: Randall Flagg on 21 Feb 2010 18:02 Hi everyone, I want to use a chi squared two sample test to compare two images. One of them is the original (expected) picture and I want to obtain a measure of how good the other image fits it. I tried corrcoef but it's not robust and sensitive to outliers. I also tried spearmans rank correlation coefficient but it has problems with homogeneous image background regions. Can I use a chi2 test for that purpose? And how can I do that? chi2gof can only test one sample against a normal distribution, how can I test two samples against each other? Thanks a lot.
From: ImageAnalyst on 21 Feb 2010 19:49 Why not use PSNR like most people?
From: Randall Flagg on 22 Feb 2010 03:37 ImageAnalyst <imageanalyst(a)mailinator.com> wrote in message <f793e121-b5e3-4963-8468-9e321c019807(a)d2g2000yqa.googlegroups.com>... > Why not use PSNR like most people? Because it depends on scaling: PSNR(A,B) ~= PSNR(42 * A, B) The second image is not a corrupted version of the first one, but a reconstuction from other data. I need a measure like the correlation coefficient, that takes B being a linear transformation of A into account, just more robust.
From: ImageAnalyst on 22 Feb 2010 08:11 And chi square wouldn't depend on scaling???? Anyway, I would think intensity scaling is something you'd want to know about.
From: Peter Perkins on 22 Feb 2010 09:55
On 2/21/2010 6:02 PM, Randall Flagg wrote: > Can I use a chi2 test for that purpose? And how can I do that? chi2gof > can only test one sample against a normal distribution, how can I test > two samples against each other? Randall, I don't know much about image processing, or if a chi-squared test is approporate in your context or not, but CHI2GOF certainly can do more than test against a normal distribution, and has a provision for testing against a distribution that has been estimated: >> help chi2gof CHI2GOF Chi-square goodness-of-fit test. [snip] The following options determine the null distribution for the test. You should not specify both 'cdf' and 'expected'. Name Value 'cdf' A fully specified cumulative distribution function. This can be a ProbDist object, a function handle, or a function. name. The function must take X values as its only argument. Alternately, you may provide a cell array whose first element is a function name or handle, and whose later elements are parameter values, one per cell. The function must take X values as its first argument, and other parameters as later arguments. 'expected' A vector with one element per bin specifying the expected counts for each bin. 'nparams' The number of estimated parameters; used to adjust the degrees of freedom to be NBINS-1-NPARAMS, where NBINS is the number of bins. |