Fast and dirty face recognition and database storage / search in Java

Question

Last week I studied and experimented with face recognition. The planned application is designed so that a person can find information about a person in a database (SQL) simply by taking a picture of his face. The initial expectation was to compress the face to a key or hash and use it as a lokup database. This does not have to be extremely accurate, as the person viewing the information can and most likely will complete the final comparison of the original image with the file and the person in front of them.

OpenCV / JavaCV seems to be an obvious starting point, and the face detection it provides works well, however the Eigenfaces implementation for face recognition is not ideal because online training will recompile hundreds of thousands of users every time a new face needs to be added to the training kit will not work.

I am experimenting using SURF descriptors on a face extracted using OpenCV Haar Cascade functions, and this seems to bring me closer to the intended result, however I cannot think of a way to efficiently search and compare about 30 descriptors (which are 64 or 128 dimensional vectors ) in the database. I read some LSH and Spectral Hashing algorithms, however there are no implementations for Java, and my math is not strong enough to implement them.

Does anyone have any thoughts or ideas on how this can be achieved, or if it is possible?

Answer 1

Hashing is not complicated, and you do not need a degree in mathematics.

Assuming that any 2 images result in a fairly similar number of “descriptors”, then it only requires that you get a reasonable match with enough of them to get a sufficiently high confidence coefficient.

How specific these descriptors determine what level of collision you can accept in your hashing algorithm.

Since you have several of them, I would suggest that you don’t need anything too complicated - because you probably want to get a level of “fuzziness” in your search?

Start with something simple - experiment and refine. You might even find that you need a different hash for different descriptors - that is, some may be more specific than others?

Hope some food for thought.