Enhancing writer identification and writer retrieval with CenSurE and Vision Transformers

Abstract

The object of research is the process of writer identification based on handwritten text. Despite significant progress, existing methods for author identification from handwritten text have limitations that prevent them from achieving maximum accuracy and reliability.This paper focuses on optimizing and improving the efficiency of writer identification from handwritten text by integrating image preprocessing methods, feature detection, and modern machine learning architectures. To this end, a functional model was developed that uses the CenSurE algorithm to detect key points and extract relevant image areas, and then the Vision Transformer model to identify the writer based on these extracted features. To reduce the variability of the results, experimental validation was conducted using a dual search and classification methodology. The use of the public CVL dataset increases reproducibility and helps in comparative analysis. The findings indicate that the implementation of the proposed approach leads to an improvement in the identification accuracy during retrieval, surpassing the results of other studies. This is evidenced by increased accuracy values of hard top k and soft top k by 1% and mean average precision by 2%. In addition, findings indicate significant performance improvement in the feature detection preprocessing stage. This improvement is quantitatively supported by reductions in both the average time per item and total processing duration by 39%, alongside the increase in total count of patches extracted by 70%.The results obtained contribute to increasing the reliability of automated handwriting analysis systems, especially for the task of writer identification. This achievement is a valuable tool for graphologists and forensic document experts, supporting such critical tasks as the forensic authorship process.