CNN Signature Analysis POC

Deep Learning Proof of Concept for Analysing Authenticity of Signatures using Tensorflow CNN (https://en.wikipedia.org/wiki/Convolutional_neural_network).

Getting Started

Installation

To install required packages, run pip install -r requirements.txt.

Training / Testing

• Data will be read from the dev_data folder. Place images within dev_data (under train/test, forge/real) to train & test the model based on input segregated data.
• To train and test the model, run python train_test.py. This will generate a model folder storing the model information
• To increase the number of training steps, edit the N_STEPS variable under train_test.py.

Prediction

• Data files will be read under predict_data folder.
• To run predictions on images under the above folder, run python predict.py. Predictions will show on stdout.

Notes

• This POC mainly tests out a sample signatures for 2 people. Further upgrade possible, to segregate model folders, each customized to a particular person. Each model folder will be tagged with an ID, with binary classification of forge/real for a particular person.

Acknowledgements

Images used for train/test/predict actions in this POC comes from http://www.iapr-tc11.org/mediawiki/index.php/ICDAR_2011_Signature_Verification_Competition_(SigComp2011).

Built With

• Tensorflow - Open Source Machine / Deep Learning Platform
• OpenCV - Open Source Computer Vision & Machine Learning Library
• Python - Python 3.6.2