Research on interpretability of writer identification based on multi-label classification of writing style

Abstract

Writer identification based on deep learning has shown great potential in fields such as forensic analysis and financial security due to its high efficiency and accuracy. However, the specificity of deep neural networks limits the acceptance and adoption of their identification results in these fields.This is due to the ‘opacity’ of deep neural networks. To address this issues, this paper proposes an interpretable framework for writer identification based on multi-label classification of writing styles, implemented using residual networks and attention mechanisms. Firstly, this study selects five writing style types commonly used based on the experience of manual writer identification.Based on the Chinese handwriting dataset HWDB2.0, multi-label writing style annotation was carried out to construct the writing style dataset HWDB-STYLE. Next, a residual convolutional neural network combined with a channel-spatial attention module is used to construct the backbone network. Finally, the number and structure of the classifiers in the backbone network are improved and a multi-task model is obtained which performs multi-label classification of both writer and writing styles. This model can provide both the identity of the writer and the classification of five different writing style types, and interpret the output writer identity results through the output style type. Experiments on the HWDB-STYLE dataset demonstrate that the model not only maintains high accuracy in writer identification but also accurately classifies the writing style of each sample. The results are consistent with human observations, providing a level of interpretability for writer identification results.