I've developed artificial intelligence / computer vision techniques used to make decision on image similarity using some tools I've developed for high-energy physics. The algorithm is rather fast: the analysis of an image with 120x150 pixels requires about 1-3 sec. The output of this algorithm produces 3 probabilities: 1) Similarity for color intensity (easy part!); 2) Similarity probability for object locations; 3) Similarity probability for object sizes (shape matching); Yes, the algorithm really looks inside an image and can resolve objects in it! If the analyzed image is close to the sample image, all three probabilities are close to unity.
The algorithm does not use pixel fingerprints, so the method can perform image identification even if the original image was re-mastered. The algorithm can be used in many areas such as: 1) search for a similar image; 2) detection of copyrighted images and videos on the Internet. 3) search for similar images on your PC; 4) detection of object changes in an image (object locations or sizes); 5) biometric identification; 6) computer and robot vision.
Here are several examples:
1) simple image with objects
2) very complicated image
Note: the algorithm was not tuned for a particular image.
This algorithm has nothing to do with image understanding, i.e. it cannot make a distinction between a human and a car. But, I think, it can be extended to do this.
Interested?
S.Chekanov (DESY/ANL)
