How to determine if a photo is a poster (not realistic)?

Question

I have a number of .jpeg from musician sites. These images consist of posters for upcoming shows and group photos (group photos in real life).

Here is an example poster:

I don't understand any modern methods or algorithms (if they exist?), But this is what I thought I could look for:

• The text in the image is usually the dead side of the poster.
• Perhaps realistic photographs (i.e. without posters) follow a different distribution of colors?
• Posters are probably less likely to have faces in them, but this is a rather weak statement.

Is there any classification algorithm that can determine if an image is a poster?

Answer 1

Your question is very broad. The poster or photo is not a clearly defined object. What is a poster? In real life, posters are often photographs or a combination of photographs or slightly corrected photographs.

If we narrow down to refer to the first part of your question - group photos and upcoming posters, then the answer is probably yes (although I have never seen anyone do this). Since you are looking for a binary classifier, I would suggest taking some machine learning model (Naive Bayes should be enough, but if you want to use more complex functions, try SVM, ELM or some random forests / decision tree) and apply them to the data encoded in vectors containing:

Binary Functions:

• "is there any word in the image?" - you will need an algorithm for detecting external text.
• "is there a number in the image" - events must have dates
• "there is a date on the image"
• "is there any face in the image"

Using Naive Bayes would create the conditional possibilities P(poster|there is a word) , P(poster|there is a number) , etc., which not only give you a classifier, but also some information about how important your characters are (a probability close to 0.5 is a suggestion that a particular feature is useless).

I would not use histograms, etc. because of the wide range of possible photos, photo shoot styles, etc., if you do not want to create a really large set of workouts.

If this is not enough, you can change them to more complex functions and use a more powerful classifier than Naive Bayes.

Complex features:

• How many words are there in the image?
• How many numbers are there in the image?
• How many dates are in the image?
• How many faces in the image?
• Image histogram

And one last option, if all else fails, you can try to prepare some modern model, such as the Deep Belief Network, on raw images. This will require serious computational power, but the results will be very useful for the scientific community.