Counterfeit detection of ID documents from a mobile device photography

This model detects whether an ID Document image, acquired from a mobile device, comes from a genuine document or from one that is likely to be forged. Document contours are detected, and the ID Document is cropped and rectified, preserving the scale (size) of the original ID Document image. Once the image of the document has been rectified, it is divided into squared overlapped patches that overlap 50%. Each of these regions feeds into a many-to-one recurrent network model and uses a pre-trained convolutional neural network model as a backbone. The default backbone is a ResNet50 pre-trained in Imagenet.

Then, the penultimate layer feeds a recurrent cell in length the number of regions in the images. Recurring units can be from the most basic RNN, LSTM, and GRU. Experimentally, it has been seen that the best results are obtained with an LSTM. The last vector of the recurring network is used as the descriptor of the document. Finally, these vectors are used to train a few-shot model called the Prototypical Network (PN). The weights of the recurrent network and the MLP network of the PN are adjusted simultaneously in the meta-training phase.