Skip to content

Visualisation exploring handwriting styles using convolutional autoencoders and clustering

Notifications You must be signed in to change notification settings

annahadji/handwriting-styles

Repository files navigation

Exploring similarities in handwriting styles

During National Inclusion Week, we collected samples of handwriting styles across the Institute of the phrase Each One Reach One, the theme of the week. We were interested in exploring and visualising different styles of handwriting that might exist.

We utilised encodings from a convolutional autoencoder trained on the data to compute a representation of each sample that might be indicative of style. We then used dimensionality reduction (UMAP and t-SNE) and clustering algorithms to explore and visualise similarities that exist amongst the writing samples. We found that some samples did share similarities in style such as curliness or thickness of writing. More details can be found on the Alan Turing Institute's blog.

Built With

  • TensorFlow - data processing and model.
  • Scikit-learn - dimensionality reduction and clustering.
  • Bokeh - interactive scatterplot visualisation.
  • Jinja2 - injecting Bokeh visualisation into standalone html template.