The program was created last spring and I have just been sitting on it – no time. You create a collection of images and PxSearch can search for pictures similar to the one that you choose.
This is not even a prototype; it is in fact a test program. This is why every time you add an image to the collection, a few of its versions (rotated, blurred, etc) are also added and analyzed. This feature is needed to help you see when the algorithm works well and when it does not. The idea is that these versions of the image have to appear near the top of the list of similar images. It works reasonably well, with some images. The appropriate image have to be of good quality, with several larger objects, little pixelation, noise etc. Faces, simple landscapes, some medical images work. Fingerprints don’t. But I think many more would work if the thumbnails were larger. Computation isn’t fast in the first place and then analyzing extra 7 versions of the image takes extra time. So, unfortunately, I had to shrink the images to 100×100 to make the processing time reasonable.
The program is available upon request. Keep in mind that you will need at least a few dozen images in the collection to appreciate the algorithm. But to seriously test it, you’d need at least thousands. This is why I want to put it online as soon as possible.
The algorithm is very unsophisticated. It is based on Pixcavator and simply computes distributions of objects of each size. That’s what the “signatures” are. To match two images, their signatures are compared by a simple formula (weighted sum of differences). The end result is not a search based on “likeness” (it’s been done, badly), but on a quantifiable similarity. Basically, it’s about finding copies of a given image. This may have something to do with copyright filtering. More information will be added to the article on Image Search in the wiki. There are a few visual image search engine listed in that article. They didn’t impress me technologically, but those that work are surely fun to use. So is this one.
I was answering to an e-mail about where the main algorithm of ours has come from and how it is related to image segmentation especially the watershed algorithm. So I decided to quote it here. Books on image processing, like Gonzalez and Woods, spend very little time on topological issues and totally ignore well developed topological approaches and techniques. There is a good reason for that (and some bad reasons too). Traditionally topology can only be applied to geometric objects, i.e., binary images. This is very limiting because if you deal with gray scale images you have to start with thresholding every time and every time you loose information. When I discovered this, I decided to try to fix the problem by developing an algorithm for gray scale and color images based entirely on algebraic topological methods. The result is the method that we use here. It is similar to the watershed algorithm but only in the sense that both look for minima and maxima of the gray scale function. The difference comes from this simple topological idea that in the image there are objects and holes in them, and there is nothing else. In case of gray scale, this means that there are dark objects on light background with light holes in them, or vice versa. So, what the algorithm produces cannot even be called a “segmentation” which is supposed to split the image into non-overlapping areas.
Last week
· A short page was added under “For researchers” – How it works. It gives an initial idea of how and why Pixcavator works with a link to the wiki. Hopefully, there will be a link to a paper here in the near future.
· The article on Binary Images is mostly finished, pending feedback. I’ve made good progress with Grayscale Images too.
· I started to add answers to exercises. The wiki is starting to take a shape of an online textbook. Interesting…
· A PowerPoint slide show was added that explains the basics of our approach. It will be updated regularly until it takes the form of a paper.
· We’ve had technical difficulties with the blog…
Stupid wordpress is bugged. It cut off almost the whole list of posts. I don’t have time for this.
To get to the rest, you’d have to click on a post and then go to the categories…
I was reading Steve on Image Processing blog (image analysis with MATLAB) - two consecutive posts Gray scale pixel values in labeled regions and Intensity-weighted centroids, with goals totally unrelated. Yet, both implementations begin the same way – thresholding of the image. I started to remember that this is very typical in image analysis. In all the MATLAB tutorials I’ve seen, especially tutorial intended for beginners, whatever you do, you have to start with thresholding. It seems strange because there are many sophisticated algorithms for grayscale and color images. Except, those are about image processing, while almost all image analysis algorithms are for binary images. ImageJ is another good example. Among a hundred of image manipulation/processing commands, there are a few image analysis tasks tucked in. And of those, what they call image analysis is either all manual or isn’t about analysis of the content of the image in the first place (like the color histogram). A couple of commands left are only for binary images. Once again, if you try to “Analyze Particles”, you get the message “Threshold first!”… I actually find this very encouraging because it means that what I am working on has some value… [A smiley face used to be here but apparently it messes up WordPress…]
I’ve talked to quite a few people about Pixcavator SDK but this is the first tangible result. Ash Pahwa of the University of California - Santa Barbara has created a prototype software called CellCounter (not the final name probably). The program mostly follows Pixcavator for now but the idea is to develop it in such a way that it would work for some microscopy tasks especially cell counting. Dr. Pahwa works with biologists on various approaches to this problem. To be accepted in this community the methods we use will have to be better explained and verified. CellCounter works on Vista – for XP you’ll need to install .NET 3.0 Framework from the Microsoft web site (takes 5 minutes). It looks beautiful and all it took was just a few days!
