There have been discussions about how to compare two images, “calculate” the difference between them. This can be seen as a variant of image comparison, but not entirely.
Consider this:
Suppose I’ve the image of a pen, stored in the SD card of my phone (Android phone). And, I’m “searching” for this pen in my house, using my camera. Let’s ignore the fact that I can do the same with my eyes, and let the camera do the searching. So, I’m in my house, pointing my camera at various places, to see if my camera can “identify” this pen, when I point the camera at it.
To put in vague computer science terms, while I’m pointing my camera at stuff, I have a thread running in the background, to continuously compare the image of this pen, with the current Camera View. I point my camera at a chair, the thread compares the camera view of the chair with the image of the pen , and since the chair is not the pen, it returns “false”.
And when I finally point it at the same pen, lying on a table, the thread should determine that this is the pen whose image we have on our SD card. So the Camera View now shows the pen as well as the table it is lying on, but it should identify the pen using image comparison techniques.
Is something like this possible, in general? Forget Android , or a smartphone , is “identifying” an area or object with the camera possible? Being ignorant of image processing libraries, I can only assume that some library does have tools to do such stuff. Or atleast, there could be a “theoretical” algorithm, to start working on.
Thanks,
Sanjeev Mk
This kind of thing can be done, but to get it working well is very, very, tough. My experience of OpenCV processing on Android is that it is out of the question to do this in real-time with current mobile hardware. ( It might be feasible to do this with still frames )
A very crude ( And by a country mile the easiest ) system would use template matching
cvMatchTemplate. Essentially, you compare the “template” ( a photo of your pen, for example ) to the photo at each possible position in the search image. It is computationally intensive but works quite well under constrained circumstances, your application requires completely unconstrained conditions, however.My suggestion would be to look at Surf or something similar and the Hough transform. You make a “finger print” of an object by producing a set of SURF features on the object from your reference image. You run surf on the search image. Then you apply a generalised hough transform where the object model is a set of feature points. Peaks in Hough space represent good matches.
I’ve never even tried the second approach but I know it is possible. Also the two approaches I’ve put forward are by no means the only two, they’re just two which are familiar to me.