I have to classify different parts of an image as natural (obtained from camera) and synthetic (computer generated graphics). However, given the nature of images I am dealing with, the natural images are also displayed on a computer screen, making it difficult to distinguish them from the synthetic images. How do I proceed? Any help is appreciated. Thank you.
I have attached a sample image, wherein the natural images have been surrounded by yellow bounding boxes, and the synthetic images have been surrounded by green bounding boxes.
Well. from your sample image which is apparently a screen shot from a TV screen, it is easy to detect the rectangles that contain text .Just us a character recognition algorithm. For detecting segments of an image that contain faces one might try to detect natural faces by a skin color detector using a gamut of face skin colors.
For natural scenes like landscapes, one might use an edge detector to examine how smooth the edges are. If they composed of linear segments they might be
synthetic, of they are smooth its an indication they are natural.
Another approach , is determine a number of image categories, construct a training set of real and synthetic samples and train a classifier to discriminate between them.
have a look on this patent http://www.google.com/patents/WO2003081533A2?cl=en
Well, you can try a little but effective trick which is based on JPEG compression . As JPEG is designed, striktly speaking, for natural scenes, the inter-correlation between pixels is high (Andrews 1966). Artificial images which have usualy high contrast edges generate artefacts (noise) when jpeged. Build a grid over the image, compute the noise level per region, design a clever threshold, then you are able to distinguish (per region) between natural and artificial images.
Calculating the number of colors used.
Natural images will generally have a number of colors that a lot of
Simpler than the gradient method in the patent mentioned by Emad Monier Ibrahim, would be a method that does an edge detection of the RI ( region of interest) of the image. After count the number of edges and there length. Those with many small edges correspond to natural images while those with a large number of long edges, above a certain dual threshold, belong to a synthetic image. You should have some kind of prefilter to segment the original image in RI, otherwise the results will be based on a mix of natural and synthetic. I think this method will be less complex (computational wise) than some other methods
Dear Kumar,
From the description and the sample image, your task seems very hard. First of all, if you imagine that the camera actually filmed a scene which contained letters, the aforementioned solutions (edge detection, text detection) would not work, since they would wrongly label the camera image as synthetic.
If you can somehow constrain the types of images that you want to label as synthetic / natural, then you can construct a much more powerful framework (based on typical features that occur in either of your classes).
In general, the approach mentioned by Helman in the first answer (construct positive and negative training examples, and train a suitably powerful classifier) may work when the classifier is able to pick up the correct features. You may want to take a look at typical image descriptors people have used for discrimination. In addition, having a rich feature set and combining the results of multiple features (think: color, edges, texture, motion, etc.) will almost always make your system more robust.
On a side-note, there is excellent research on forgery detection and image content done at Dartmouth College by Hany Farid:
http://www.cs.dartmouth.edu/farid/
Best,
Christian
- Badges
- Science topic
- Similar topics
- Psychology
- Mental Processes
- Learning