Ensemble of transfer learning with convolutional neural networks for writer recognition in historical documents

Abstract

In the cultural heritage domain, writer recognition has become a challenging classification task still explored for historical documents, due to the presence of different types of noise in the documents, i.e., ink bleed-through, ink corrosion, stains on paper or parchment, difficulty in the character discrimination, elements different from the text, such as images, etc. that limit the effectiveness of existing techniques. To further advance in terms of robustness of classification and experimental setting, we propose a new deep learning model which ensembles pre-trained convolutional neural networks for writer recognition. Specifically, the ensemble is composed of three pre-trained Inception-ResNet-v2 models with different hyperparameter values. Results obtained on the benchmark ICDAR 2019 dataset of handwritten historical documents prove that the proposed approach is very promising in recognising the handwritten characters of different writers, also when compared with other deep learning models.