Handwritten documents author verification based on the Siamese network

Abstract. The paper presents a method for verifying the handwriting of a certain author in a corpus of handwritten documents based on a small number of examples, and proposes an algorithm for data preprocessing.
The use of Siamese neural network is proposed to compare and analyze unique characteristics of handwriting, writing style. This way of training allows to obtain powerful discriminative image features, embeddings, on the basis of which it is possible to make a qualitative classification of the author.
The proposed approach was applied to the task of verification of possible autographs of Zhukovsky among manuscripts of unknown authors. The approach was also applied to the classification task on a fully labeled IAM dataset.