-
A convolutional autoencoder is created and trained on images of person with dimension 28x28.
-
Architecture was used from Keras Convolitional Autoencoder.
-
The network is trained end to end using only positive class samples.
-
To check whether a test sample belong to positive class or not, it is passed through the trained network and cosine similarity is calculated between the input and predicted output.
-
The output for cosine similarity should be high for the positive class test sample and should be low for negative class sample.