How to compare the feature vectors of images in two separate folders

Hi my project deals with the comparison of the feature vectors of images residing in two separate folders. How to compare them ?

Answer 1

[Walter Roberson]

http://matlab.wikia.com/wiki/FAQ#How_can_I_process_a_sequence_of_files.3F

Generally when you have the feature vectors for two different images, you find a "distance" between them; using Euclidean distance is not uncommon.

[Monisha]

So I need to find the euclidean distance between the images for comparison?

[Walter Roberson]

No, you could find the Minkowski distance or the Manhattan distance or the infinity norm or the bhattacharyya distance or you could fit cubic polynomials to the feature vectors and consider the difference between the cubic term, or you could project onto a hypersphere and find the hyperbolic distance. Invent your own scheme, but you should probably make sure that it is a metric

[Image Analyst]

If you're using a distance metric, then take some though about how you want to weight the various features. You may not want to weight them all the same. Maybe you want to normalize each of them to the range 0-1 first, and even then apply some weights after that.

[Walter Roberson]

Right. If a change of 1 in feature A represents a major change, but a change of 1 in feature B represents a tiny tiny change compared to what is being measured, then you want the "1" of A to have more weight (contribute more to the "distance") than the "1" of B.

[Monisha]

yeah. I am using a distance metric. And moreover i am storing the feature vectors of my images in an access database. So can i directly retrieve the images from the db and then compare them in matlab ?

[Walter Roberson]

Okay, so is there a remaining question?

[Monisha]

Yeah. Sorry for the ambiguity. So can i directly retrieve the images from the access db and then compare them in matlab.

[Walter Roberson]

Yes, but you would be better off pre-calculating the feature vectors and storing them, and then comparing feature vectors.

[Monisha]

Yeah i have already calculated the feature vectors of the images and stored them in a database. So is it better to compare them directly without finding the euclidean or manhattan distances ?

[Image Analyst]

It's better to do whatever works best . I would think the weighted Euclidean distance is what I'd try first. I'm not sure how you would "compare them directly" if that doesn't mean calculating distances . Are you talking about computing maximums, median, or minimums or something? Again, do whatever you want and whatever works best for you.

[Walter Roberson]

You can compare them directly using isequal(). But all that is going to tell you is whether they are exactly the same or not. Feature vector approaches allow you to determine whether they are similar or not.

[Monisha]

Image Analyst:Okay. I will try comparing the features directly. Walter Roberson: Yeah . It is similar to say whether the images are same or not.

[Walter Roberson]

isequal() would only be for the case where there was not even a single bit different in the two images.

In the situation where you are comparing images to determine whether they are identical, a useful technique is to calculate a hash or checksum of each image, such as MD-5. Then two images cannot be identical if they have different hashes (or checksums) Once you have narrowed down to a smaller set of candidates by excluding those with different hashes (or checksums), then you can afford more time to isequal().

[Monisha]

Ya. Thank you for the method. i will try with the checksum.