SigRec is a microservice designed as an end-to-end service which can be hosted independently to facilitate the workflow around signature recognition and analysis.

Bootstrapped in Axis Bank AI Challenge - The Machine Learning Model scored the highest accuracy throughout the competition and was selected among top 10 solutions across India.

Displayed in Synd-Innovate Hackathon - The solution was ranked as the 4th best innovation in the entire competition.

Demo

Demo from a sample dataset

svg

Genuine Prediction

Smart features like minimum rectangle box, Bounding box to closest polygon, Convex Hull, Aspect Ratio, Keypoints, Descriptors and many more features are extracted from a signature which is used for analysis and verification process.

Forged Prediction

Along with the above smart features - we also use Deep Learning techniques to create a 2048 feature vector which is in itself goes through dimensionality reduction techniques to strengthen the accuracy of the model.


Model Processing

Lets look at the steps that each signature roughly goes throughout the pipelines for final inference.

Data Cleaning

Each signature goes through practices like noise reduction, normalizing dimensions, gray-scaling, smoothening, etc.

Data Augmentation

In the next pipeline, the image is cropped and central positioned while augmenting in such a way that the change in angle of signature doesn't affect the model evaluation process.

Smart Feature Extraction

Features like Contour Approximation, Convex Hull, Keypoints, Descriptors, Aspect Ratio and many other key attributes of the signature are retrieved.

Deep Learning

Higher Dimensional (2048 dimensions long) feature vectors are computed using CNN model to extract features that are too minute to be retrieved by formal approach.

Feature Selection

All of the features goes through dimensionality reduction using Principal Component Analysis and Standard Scaling to prepare the final feature vectors ready for evaluation.

Prediction

Lastly a SVC Classifier already loaded with trained weights evaluates the final feature vector to predict the authenticity of the signature.

The work on SigRec begin on early November 2018 preparing for Axis Bank AI Hackathon - Signature Recognition and Verification Challenge.

The competition was divided into 3 phases - Ideation, Prototype and Demonstration.
The first two rounds were online based while the last round required Top 40 teams to group at Thought Factory, Bangalore to showcase their solutions to judges and test it on a Live custom dataset prepared by the organisers using their own signatures real-time.

SigRec model scored the highest accuracy among all of the participants in the live dataset and was recognized among the Top 5 models in the entire competition.

Sigrec also won the 4th position at Synd-Innovate hackathon organized by Syndicate Bank as one of the best innovations in the competition.

Hackathon
Stats

Total Participants

2650

Most Accurate Model Rank

1st

Presentation Rank

Top 5

Social Media Coverage