8 Document forgery detection using computer vision
This chapter covers
- Understanding document forgery
- Analyzing a receipt forgery dataset
- Using CNN (Convolutional Neural Network) for digital forgery detection
- Posing forgery detection as a segmentation task
- Generating auxiliary labels using an expert model
- Enhancing forgery detection performance using multi-task learning
As a mischievous kid, I remember trying to forge my father’s signatures on a letter requesting a leave of absence from my school. Of course, my class teacher immediately caught me. Thankfully, I learned my lesson. Unfortunately, some adults try the same, to illegally withdraw someone else’s money via forged signatures for example. This is a prime example of document fraud.
8.1 Exploring receipt forgery dataset
8.2 Posing document forgery as a segmentation task
8.3 Training CNN model for document forgery detection
8.3.1 Processing receipt forgery dataset for PyTorch
8.3.2 Defining deeplabv3-based segmentation model
8.3.3 Training segmentation model for forgery detection
8.3.4 Evaluating trained forgery detection model on test set
8.3.5 Interpreting predictions from the forgery detection model
8.4 Improve model performance using multi-task learning
8.4.1 Building an auxiliary task for better model training
8.4.2 Generating auxiliary task labels using an expert model
8.4.3 Training a multi-task CNN model
8.4.4 Evaluation of multi-task model performance on the forgery test set
8.4.5 Interpreting multi-task model predictions
8.5 Summary