Deep Learning-based Forensic Writer Retrieval/Document Segmentation

Status: available
Supervisor: Manuel Keglevic


Law enforcement agencies possess an extensive collection of handwritten documents. This includes for example documents belonging to open cases, for example death threats, and reference samples from suspects and prisoners. Yet, these collections of documents can only be utilized to a limited extend, since for an identification of an unknown writer all documents have to be compared manually by handwriting experts.

In course of the WRITE project we are developing an IT assisted search and comparison of handwritings using a Deep Learning-based methodology to expedite the identification of unknown writers by handwriting experts.

By doing your Praktikum, Bachelor Thesis or Master Thesis on this topic you get the opportunity to work on unique data and help the Police do CSI stuff like in those TV shows (unfortunately you have to bring your on sunglasses).


There are two parts in our methodology where we need your help with. Firstly, the provided documents have to be segmented into handwritten text, printed text, signatures, and so on and secondly the handwriting samples need to be compared to compute a similarity score. Depending on your interests we will define a topic that’s tailored to you.

You are required to use Python and PyTorch since this helps us incorporate this in our methodology, but you don’t need to have experience with Python – yet. It’s definitely a good opportunity learn python or to improve your Python skills!

As a starting point, we will point you to literature you need to read and of course help you at every step of your way.


  • Literature Review – getting to know the methods
  • Implementation & tinkering
  • Evaluation
  • Written Report/Thesis and final presentation

Helpful experience

  • Python and numpy
  • Computer Vision applications and frameworks
  • Machine Learning frameworks like PyTorch or Tensorflow