mapo80/SignatureVerificationSdk

Sdk to verify signature

Why this project

I want a method to verify if an handwritten signature is genuine or forged. I've googled a lot and found some papers about this topic. I've found many papers and finally I've found these two projects on Kaggle:

• Signature Image Cleaning with Tensorflow 2.0 - https://www.kaggle.com/code/victordibia/signature-image-cleaning-with-tensorflow-2-0
• Siamese signature verification with confidence - https://www.kaggle.com/code/medali1992/siamese-signature-verification-with-confidence

I've created another repo on Github to use output models from previous Kaggle projects and convert them to ONNX format:

• https://github.com/mapo80/signature-verification-models

I've converted these models to ONNX format, so to use them in this c# project.

How to use it

To clean signature images

• Use the SignatureImageCleaner class to clean signature images. There are two methods that accept stream or byte array as input. Return images have 224x224 size. Below some examples:

First example

Second example

To verify signatures

• Use the SignatureVerifier class to verify signatures. There's one method to verify two signatures. VerifySignatures(Stream sourceImage1, Stream sourceImage2)

It doesn't use previous method to clean signature image but it uses a different one, same used on kaggle notebook. Preprocessing filters are:

• GaussianBlur
• Otsu threshold and binarized image to fine center of mass and minimun area to contain signature
• Center image on canvas size
• Normalize image and remove all noise using otsu threshold
• Resize image and crop to center

These preprocessing methods are same used on original system. Here are more details:

• https://github.com/luizgh/sigver

Final results

I've tested this project with all test signatures provided on kaggle notebook dataset:

• https://www.kaggle.com/code/medali1992/siamese-signature-verification-with-confidence/input

All signatures are in test folder.

Inside test_data_results.csv are all results of this project. Info contained are:

• Image1 and Image2 are the two signatures used to verify
• IsForged If image is forged or not
• CalculatedResult calculated result from siamese network
• IsSameResult if it's same result as dataset
• Confidence is the confidence of verification calculated by model
• Similarity is the similarity of verification calculated by model

To determine if a signature is forged or not has been used Similarity, if this value is < 0.5 then is forged.

Results

Below there are results:

• Accuracy: 0.6399921352732992
• Precision: 1
• Recall: 0.30957767722473606
• F1-Score: 0.47279009501871583

This model is not perfect, but it's a good start to verify signatures.