Recurrent Few-Shot model for Document Verification

TL;DR

A recurrent-based model able to detect forged documents in a few-shot scenario is proposed, which makes the model robust to document resolution variability and the few-shot approach allow the model to perform well even for unseen class of documents.

Abstract

General-purpose ID, or travel, document image- and video-based verification systems have yet to achieve good enough performance to be considered a solved problem. There are several factors that negatively impact their performance, including low-resolution images and videos and a lack of sufficient data to train the models. This task is particularly challenging when dealing with unseen class of ID, or travel, documents. In this paper we address this task by proposing a recurrent-based model able to detect forged documents in a few-shot scenario. The recurrent architecture makes the model robust to document resolution variability. Moreover, the few-shot approach allow the model to perform well even for unseen class of documents. Preliminary results on the SIDTD and Findit datasets show good performance of this model for this task.

Figures (5)

• Figure 1: Data modelling architecture: The sequence of patches are extracted for each document. Each patch is then fed into a CNN feature extractor. The vectors from the CNN are passed to a RU at time $t_k$ depending of the patch position. The overall vector of the document is extracted by the RU at $t_K$
• Figure 2: Example of conditioned prototypical classification with k-shot=3
• Figure 3: Example of Unconditioned Prototypical classification with k-shot=10
• Figure 4: Example of cropped images from the SIDTD clips. (a) corresponds to a genuine sample of a synthetic Spanish ID document and (b) to a fake sample of the same nationality.
• Figure 5: Example of forgeries on prices, left image is a genuine receipt and the right image is altered from the FindIT dataset.