Period
2022.7-2023.7
Cooperative unit
CMBCHINA
Introduction
In realistic banking business, all kinds of approvals often need to be reviewed manually item by item. And the manual review mainly has the following defects: (1) the manual review is time-consuming and laborious in massive business scenarios; (2) the pseudo-samples are sparse in realistic scenarios and the long time manual review mechanism is prone to leakage and mis-detection. Some current counterfeit detection solutions have single detection means, insufficient accuracy, and are out of touch with reality, and no mature solution has been deployed in actual scenarios. Therefore, this project plans to develop a few-shot learning forgery detection method for forgery seals with the objective of an intelligent system for detecting forgery seals of forgery documents for banking business.
Methodology
There are 3 main subtasks to be completed in this project: forgery detection, fake seal detection and handwriting style signature generation.
1. Forgery detection：
This task mainly uses a classification model to distinguish real documents from various types of forged documents.
In this classification task, due to the sparsity of the pseudo-evidence pieces, their training sets exhibit severe long-tail distribution characteristics, so this project uses a meta-learning loss weight adjustment method based on meta-learning to balance each category in order to avoid the low recall ratio.
2. Fake seal detection：
The fake seal detection task is essentially a matching task. Inspired by the methods related to person re-identification, this project trains this detection model based on triplet loss.
3. handwriting style signature generation
In real-life scenarios, many customers have diverse signature styles, which poses a challenge to handwritten signature recognition systems. This project uses a generative adversarial network (GAN) based font migration method to generate different styles of handwritten signatures to provide a reference for handwritten signature recognition systems.