Printable View
I'm attend to make a project that related to image matching in VB.NET
For example, you can scan an animal, then all information of that animal will be shown on computer screen :) (ofcourse we will have to build a dictionary)
Therefore, I just want to know what should I study to write code for matching images together? And is there some ways that I can get the best matching correction?
Regards!
That is a very difficult task. Imagine walking around an animal taking pictures every degree so you have 360 of them. You can have closeups, long shots, crouch down or climb a ladder. Then some animals are different colors/shades ect. You can easily have 1000's of pictures of the same animal. And if you don't then the less you have the more chances you will not have a match, or worse still, an incorrect match. Add to that how many different directions the head can be pointing and it gets worse. Then there's the background you will have to filter out. There also can be things in the foreground to take into account.
To do this task you would need probably 100 pictures of each animal and customize the matching on an animal by animal basis to see what you can do to make them match and not match the wrong ones. And you will have to test each animal with probably 20-100 pictures of different animals. And it will never be right 100%. And it will take years to accomplish, unless you had a huge team.
Oh sure! Maybe I just take animal's body :-? Such as the body of the fish? I think it something like we match 2 human face or fingerprint. Ofcourse I don't attent to make a 3D 360 degree matching :D cause its very very difficult to do and it need a big team.
So if I want to match 2D image (main body like this http://www.dictionaryoffish.com/buffalo.jpg), how can I study about this? I attend to do this to make a fish dictionary that help myself in my field :)
Sorry, I misunderstood. Someone wanted me to make a website for them where they can have people send them photos of birds and they can identify what bird it is. I thought you wanted something similar.
You can trace the outline. Use two loops, one for the x-axis and one for the y. Break when the color is white.
Oh yeah, this exactly what I want :)Quote:
Someone wanted me to make a website for them where they can have people send them photos of birds and they can identify what bird it is
A good idea, however, if you just loop and break until seeing a white color, maybe it is not correct :? Because maybe in it's body has some color, some special scab, etc... I just think so I do not really sure :? Besides, maybe I need more time to study about graphical programming cause my poor VB.NET skill :(Quote:
You can trace the outline. Use two loops, one for the x-axis and one for the y. Break when the color is white.
I think what you want is really quiet complex. Have you heard about the Penguin Recognition Project?
I agree with Milk.
That is how you start. Then you figure out how to tell the difference between some special scab, etc... You will have to test it with your pictures and figure out all the problems like the one you mentioned. And trust us, there will be plenty. That's why it is difficult. Try it, fix a problem, try it, fix a problem (and these problems can only be found out by trying it), and keep doing that until you can do a fish. Then try another one that comes with its own collection of problems.Quote:
Originally Posted by KendyNguyen
Good luck!
I'd suggest you start by investigating existing image matching algorithms (a simple google for "Image Matching Algorithms" turns up pages of info) like the one Milk posted
If you really want to come up with an algorithm of your own I'd suggest you start by looking up Hopfield Networks. This is a form of neural network that you can 'train' to recognise a bunch of images. Once trained you can pass it another image and it will return the closest match from it's library. The key point is that the image you pass in doesn't have to be an exact match with an existing picture, the network is capable of giving a best guess. Be warned though:-
1. It only works with two colours (black and white typically) but it wouldn't surprise me if a more advanced net has been designed that can cope with colours. The hopfield is very useful for text recognition but I've never tried it with pictures.
2. Programming it will actually be quite easy. Training it will take aaaaaaggges.
You should look into the API's that are available from TinEye.
This search engine does an amazing job for image matching.
This method will help you avoid your own algorithm.