Multi-scale CNN-CrossViT network for offline handwritten signature recognition and verification
Abstract Developing technologies that can accurately identify and highlight subtle differences in signatures is crucial for improving the performance of signature recognition and verification. Especially for writer independent offline handwritten signature verification technology that can directly v...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Springer
2025-07-01
|
| Series: | Complex & Intelligent Systems |
| Subjects: | |
| Online Access: | https://doi.org/10.1007/s40747-025-02011-7 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract Developing technologies that can accurately identify and highlight subtle differences in signatures is crucial for improving the performance of signature recognition and verification. Especially for writer independent offline handwritten signature verification technology that can directly verify the authenticity of the signature without the need for additional training on unknown new writers. However, with the continuous advancement of forgery technology, the differences in features between forgery signatures and genuine signatures have become increasingly subtle, which undoubtedly increases the difficulty of signature verification. To address this challenge, we introduced the cross-attention vision transformer (CrossViT) and constructed a hybrid architecture that combines convolutional neural networks (CNN) to extract stronger multi-scale features from signature images. In the CrossViT branch, depth features of different sizes image blocks are extracted, and information exchange with another branch is achieved through a token based cross-attention mechanism. This hybrid architecture fully utilizes the local feature extraction capability of CNN and the global feature capture capability of CrossViT, improving the accuracy and robustness of feature extraction. In addition, in order to make distance metric methods more applicable for writer independent signature verification in real-world scenarios, we propose an adaptive threshold method that can automatically obtain the optimal threshold from the training set and apply it to writer independent signature verification tasks. To verify the effectiveness of the proposed method, we conducted experiments on publicly available Latin, Bengali and Hindi signature datasets, including recognition, verification and cross dataset validation. The experimental results show that the proposed method achieved recognition accuracy of 98.85% and 97.40% on CEDAR and MCYT datasets using only 25% of the training data, and achieved the highest validation accuracy of 95.12% and 92.33% on Bengali and Hindi datasets, which is a significant advantage compared to other advanced methods. |
|---|---|
| ISSN: | 2199-4536 2198-6053 |