Writer Identification and Writer Retrieval Using Vision Transformer for Forensic Documents

Michael Koepf, Florian Kleber, Robert Sablatnig

Abstract: Writer identification and writer retrieval deal with the analysis of handwritten documents regarding the authorship and are used, for example, in forensic investigations. In this paper, we present a writer identification and writer retrieval method based on Vision Transformers. This is in contrast to the current state of the art, which mainly uses traditional Convolutional-Neural-Network-approaches. The evaluation of our self-attention-based and convolution-free method is done on two public datasets (CVL Database and dataset of the ICDAR 2013 Competition on Writer Identification) as well as a forensic dataset (WRITE dataset). The proposed system achieves a top-1 accuracy up to 99% (CVL) and 97% (ICDAR 2013). In addition, the impact of the used script (Latin and Greek) and the used writing style (cursive handwriting and block letters) on the recognition rate are analyzed and presented.