testing accuracy of transfomation mapping one color image to
in [Matlab]
|
From: bhavin on 4 Jul 2010 10:28 ok.. I will try PSNR. I am not so familiar with PSNR The two images i am comapring are taking at diffenrent times os they not simply a tranfomred version of the other. They would be brightness and other differences as well as noise. Would PSNR still be good?? And secondly may you refer me to a website that gives me the code to find PSNR of two images because in matlab search i failed to find such a function. Thanks
From: ImageAnalyst on 4 Jul 2010 11:07 http://en.wikipedia.org/wiki/PSNR PSNR is a metric that tells you how different two images are from one another. One common use is to compare an original image to the version of it that has been through a compression/decompression process. There are of course many other metrics that you could use or devise. Let's say that your image does have noise, brightness, and field of view differences from your baseline image. Don't you want to know that? Or would you rather have some metric that says your two images are essentially the same? If you and I were standing next to each other and we both took a photo of the Statue of Liberty, of course they're somewhat different, but they're also somewhat the same -- do you want some function to say they're the same, or different?
From: bhavin on 4 Jul 2010 11:54 I want a function to say yes they are the same image. I would like to add that the two images that i compare are of different sizes. The images are images taken from a microscope. I am viewing TB bacilli under a microscope. I tried PSNR that i got from http://www.mathworks.com/matlabcentral/fileexchange/135-psnr and testing it one my images. It performed poorly...very poorly...It gave me high PSNR values even when the two images are completely different. I guess that is because TB microscope images look quite similiar. I begin to feel thar PSNR wont work... may you recommend any other technique that is realiable and well known that i can use to find the transformation coefficinets mapping one image to its matching image and hence allow me to compare its coeffcietns to mine. ie a technique that can accurately choose matching points/control points. I came across SIFT from http://www.cs.ubc.ca/~lowe/keypoints/ . But i dont know if its one of the best. I need to compare it to a well known one and accurate technique. I am pretty sure there is one some one must have written a code in matlab that takes in two matching images are gives out the transforamtion coefficients. I just kind seem to find it.
From: ImageAnalyst on 4 Jul 2010 12:12 Yes, you can try SIFT and SURF (http://en.wikipedia.org/wiki/SURF) and similar kind of algorithms. I haven't had the need to dabble in those things (yet). You generate feature vectors and will then have to compare the feature vectors to see if enough of the features are similar enough to call the two photos a "match." You're only hinting at the larger picture (no pun intended). WHY do you need to transform (warp) one image to match another? Why are they different sizes - did they come from different microscopes? If they're images of two different slides of bacteria taken on two different microscopes, I can't think of a reason why you would want to align two such images. Indeed, since the bacteria will be different (counts, shapes, concentrations, locations, etc.) the concept of aligning ("registering") two such images doesn't even make sense. So I really don't understand the larger motivation for this, given what you've shared so far. Which brings up another good point - where are your images? You're asking questions about image processing and expecting us to figure out what to do blindly. Can you post your images somewhere, such as http://drop.io?
From: bhavin on 7 Jul 2010 05:24
Ok let me try explain fully. I have two images A and B, B is 1030x1300 pixels and A is 2060x2600 pixels. So image B is abt 4 times smaller in size than image A but the scale is the same. ie. B is a subset of image A. Both the images are from the same slide and from the same microscope but taken at different times and diffrent orientations. I already know that image B matches with image A but i need to find the similarity transformation between them, which I found with my algorithm. Now I need to check how accurate this transformation is. One way would be to compare it to a well known reliable existing technique I have uploaded a pair of images, covs4_9 and im14.10 ( http://drop.io/slideimages ). im14.10 is Image B which is the smaller one. Hope that helps clarify my question. |