Printable View
hi
i wanted to ask if anyones anything about image recognition.
e.g. i want to take the screenshot of a game board, where dice are displayed
then this code should given the bitmap recognize which number the dice is showing.
i know that neural nets can do such things. ..does anyone know where i can get a package that can do that
or maybe other ways /other packages that can achieve this goal
very thankful for any information
asta
Designing, then training a neural net for that might be harder than you think. Unless the image is always the dice in a similar orientation/size.
You're probably better off using some image processing algorithms to find the edges of the dice. Then counting the dots should be pretty easy. This all depends on how complicated the board looks and how much the orientation of the dice changes from shot to shot.
You might want to try using something like matlab if you have it.
Couldn't perceptrons also accomplish this? But yes i agree, an algorithm would be the best implementation here.
> e.g. i want to take the screenshot of a game board, where dice are displayed
Is this an actual picture of dice, taken with a video camera?
Or is it just computer drawn images of dice?
Because all you need to do is sample a few choice pixels, and you're on your way.
hi
the image is computer drawn, and the size and orientation are fixed.
i have already done it by using certain pixels.
i would like to do this a little more generally though, just incase the dice changes position, even if size and orien. stay the same.
if the software were to change the dice position even by a few pixel, my whole program would crash (or not crash, but you know what i mean).
so i was looking for something more advanced thats already been developed.
thanks for any info
asta
Fairly basic edge detection then?
As Salem says, use Edge Detection to find the edge of the dice, then use your current approach to poll pixels relative to the edge position.
Without knowing exactly what these images look like it's hard to guess how easy it is. I suspect extremely easy. Once you detect the edges of the die, take the image contrast to maximum. Now count the number of connected black blobs inside the edges you detected.Quote:
Originally Posted by Perspective
so what can you suggest to do to detect the edges?
I'm assuming that typing into a search engine "edge detection algorithms" is something you've tried already.
If you haven't, then try it.
If you think "wow, what a good idea", then perhaps programming isn't for you.
Ok, this might be a little advanced but here's how you could do it (and how I did it):
Method 1: You blur the image (Gaussian Blur), you apply the Canny Edge detector followed by the Hough transform to delineate the image and find the local maxima in the accumulator to see how many circles have been detected by Hough.
Method 2: You use a Dot/Circle Filter, you move the pattern everywhere on the image and measure the response. This is called Correlation and it's used to measure similarity.
Method 3: You use a Dot Filter and calculate the sum of square differences (ssd) and you count the minima in the image. The sum will be lowest for the dice dots.
Stuff you might need:
An overview about edge detection: http://cvlab.epfl.ch/~fua/courses/vi...otes/Edges.pdf
Delineation (Hough Transform): http://cvlab.epfl.ch/~fua/courses/vi...elineation.pdf
Texture: http://cvlab.epfl.ch/~fua/courses/vi...es/Texture.pdf
Correlation: http://cvlab.epfl.ch/~fua/courses/vi...tes/Stereo.pdf
since the dice are always in a fixed position and they are always oriented the same, it should be fairly easy to just compare the background image (without dice) witht eh new image (with dice) and pick out the delta ( the part that has changed, i.e. the dice). Get shots of each of the combinations and then feed them randomly into a simple perceptron network over a weekend, just randomly shuffle the images and do a slow train (error *.01). Use each pixel as an input into the network, with 11 hidden layer neurons and 11 output neurons. You should end up with a fairly fast and accurate NN that recognizes the result.
KONI likes researching for people ? :)
Am I the only one that is reminded of Duck Hunt for the NES and the way that it and similar games detected shots by the zapper? Ah, good times. Good times. :)
Eh, Duck Hunt for NES... Wow, shoot the white coloured pixels with your grey coloured pixels aka the gun ;)
Not really, that was one of the courses I took at University and I pretty much know all that stuff already. Just had to go to the course website and post the links.Quote:
Originally Posted by zacs7
Computer Vision is a very interesting course but damn, probably as difficult on the mathematical side as cryptography.