Photo or design recognition

Question

I need a simple algorithm is able to distinguish between design/designed picture/comic and photo.

Can you suggest one?

Do you know Aforge?

I think it could represent a solution

Answer 1

If you provided typical examples of the two cases, it would be much easier to identify possible distinguishing features. The space of all possible pictures and drawings is a tough case to start with.

Here's an approach I might try (although I'm not sure it's simple):

A. Gather a population (~30 each) of training images of both classes:

1. Designed/Comic
2. Picture

B. Try running edge detection

http://www.aforgenet.com/framework/features/edge_detectors_filters.html

on both training sets, look at the population of resulting pixels in terms of intensity. I expect that a picture will often have softer edges than a designed image. If you were to plot edge pixel intensity as a histogram and see a difference between drawings/photos, then perhaps you can also represent that as a distribution (Guassian?).

C. When you want to test a new image, you run edge detection on it and compare it's stats to the distribution you determined from the training data.

There are other qualities I might try to key off of to distinguish drawings vs. photos. For example, I expect photos would have more noise than a drawing. To use that, model the noise for both sets of training images and compare stats for new images (as in step C above). See Noise Estimation / Noise Measurement in Image .