Handwriting comparison for digital forensic analysis

  • 15 features were identified for a sample handwritten word and samples were manually annotated.
  • The task was to compare two samples and determine if they were written by the same writer.
  • A hybrid Bayesian model constructed to compare two samples gave an accuracy of 85.8%. A Siamese Convolutional Neural Network gave a training accuracy of 99.92%
    An autoencoder, along with 15 deep neural networks were also used to extract human understandable features from a sample image.

The purpose of this project is to compare handwritten samples of the letter ‘and’ of two writers and verify if the handwriting samples are of the same writer.

Initially, 15 handcrafted features were identified, and the images were annotated for these features. These 15 features of all the images were used to generate a Bayesian model and calculate the conditional probability tables. A combinational Bayesian model with 30 features and one hypothesis node was created to compare two images and generate the output (similar/not similar) at the hypothesis node.

We also create a Siamese Convolutional Neural Network to take the handcrafted features of two images and generate the prediction.

Finally, instead of having to manually annotate the images, we created a hybrid model that consists of an autoencoder combined with 15 CNNs to generate human-understandable features for the images.

Find the entire project here: