This is a fun little matching service built using Keras, TensorFlow, scikit-learn and OpenCV. OpenCV is used for Haar cascade face region identification for matching, the training data is pre-cropped faces. The cropped faces are fed into VGG16 with the dense classification layers removed. The resulting flattened convolutional vector has principal component analysis applied to reduce the space, the result of which is fed to a linear support vector machine classifier to obtain class identification.
What you do is drop some training data in /data/train
using a common name with numeric suffix for the repeats
of a given class. These will all be mapped to the
non-numeric part of the file name as a label. You can feed
a generic catch-all class like none for a wide range of
images that do not match a particular class. The reason for
doing the PCA before the classifier is to reduce the
dimensionality of the VGG16 vector to something more meaningful
to the classifier. This makes a huge difference when there
is a catch-all class (as opposed to forcing only defined
classes).
When the app first runs, it will notice a lack of the
pre-computed model and build one from the files present. Next
time it runs it will simply use the existing classifier. Both
the classifier and decomposition model are stored in the
/models directory. Just delete it if you want to refresh
the classifier model with new data.
To perform a match, you just send an image by POST to
the service. It will only attempt a face match if there is
one face detected. All other cases either lack the requesite
data or are ambiguous. In this case an error message is
presented, otherwise a match is attempted. The class is
returned in this case, which can be one of the specifically
defined classes or the catch-all class.
Here are some examples based on the data that this comes with:
curl -X POST -F 'image=@data/test/norton.jpg' http://localhost:5000/match
curl -X POST -F 'image=@data/test/hopkins.jpg' http://localhost:5000/match
curl -X POST -F 'image=@data/test/cruise.jpg' http://localhost:5000/match
curl -X POST -F 'image=@data/test/roberts.jpg' http://localhost:5000/match
curl -X POST -F 'image=@data/test/brady.jpg' http://localhost:5000/match
Given the training data, the Edward Norton and Anthony Hopkins matches
should return the proper class, whereas Tom Cruise and Julia Roberts
should return the none class.