0

My server receives photos from the client. The server is web-server in my case, but I don't mention any specific technology because I can choose any free technology that will provide me the solution. The photos are snapshots from the video streamed from the web-camera.

  1. On some photos the server receives there is a colored picture (always the same picture and the server has it in advance) on a whitish background (wall).
  2. Other photos may include any subjects on any backgrounds.
  3. I can't control the light in the room where the pictures are made (it could be darker or lighter on different photos).
  4. When there is the picture in the photo - whole the picture is included in the photo (not just part) of it.
  5. When there is the picture in the photo - it takes a very significant part of the photo (i.e. the picture will be made close to the wall).
  6. The picture on the photo could be a bit inclined/declined - let's say not more that 10°.

On the server side I should be able to decide (with a certain level of significance) whether the picture is in the photo.

I am looking for a quick and dirty solution for now (it's just POC). The library and the technology should be free.

I thought of using neutral network. In this case I could even "to train" the network off-line and once I have it tuned on I could use it on client side with javascript (the calculations shouldn't take a lot of time) without passing the photos to the server (that would be perfect).

Is there any ready solution for this problem?

Thanks a lot!

Alexander
  • 7,484
  • 4
  • 51
  • 65
  • Later, when POC is done, I should be able to decide if one of the pictures of the set (let's say 10 different pictures) is in the photo and to say which one (it will be always only one). Also I should say what is the position of it on the wall (top-left and bottom-right coordinates). Also the background will be any background (not necessary whitish wall). But for now I need a solution for the simplified problem. Of course if any solution will provide me also solution for this extended problem - I will be happier :) – Alexander Feb 02 '14 at 23:40

1 Answers1

1

I think answers to this question: Looking for an Image Comparison/Pattern Recognition Library would be a good start.

I would certainly not constrain myself to neural networks. You will need some kind of classifier but I think it would be good to start thinking about how to extract features of the images. It may turn out to be a simple problem: e.g. distinguishing between a homogeneous whitish/grayish image (a wall) and a much more heterogeneous image - that is you compute just one feature - heterogeneity - and decide based on that. In that case maybe you even wouldn't need any special image recognition library.

Community
  • 1
  • 1
Jakub Kotowski
  • 7,411
  • 29
  • 38