HanSig
is a large-scale offline Chinese handwritten signature dataset. The HanSig
dataset has the following characteristics:
- It consists of 35,400 signature samples from 238 writers (17,700 genuine signatures and an equal number of skilled forgeries).
- For each name, 20 genuine signatures and 20 corresponding forgeries were collected.
- It incorporates the real-world property of intra-writer variations by collecting signatures for a specific name in three different styles.
- The signatures cropped from the scanned images have been preprocessed by removing table lines and excess blanks around the signatures, ready for instant use.
- It is applicable to both random and skilled forgery verification tasks.
- Examples of collected signatures in three styles: neat (top), normal (middle), and stylish (bottom).
- Examples of collected genuine (top) and forged (bottom) signatures.
- Each genuine signature image has a unique filename such as original_w1_2_3.jpg. This filename is organized as follows:
- w1 refers to the first writer who signed this signature
- 2 means this signature belongs to the second name
- 3 refers to the third signature image of a specific name
- Each forged signature image has a unique filename such as forgery_w1_2_3.jpg. This filename is organized similar to that of genuine signature images.
Please fill in the form to obtain instructions for downloading the HanSig
dataset. In addition, please refer to above-mentioned Data Examples and our work for detailed description of this dataset.
If you use this dataset in your research, please cite our work:
F.-H. Huang and H.-M. Lu. Multiscale Feature Learning Using Co-Tuplet Loss for Offline Handwritten Signature Verification. arXiv preprint arXiv:2308.00428, 2023.
@misc{huang2023multiscale,
title = {Multiscale Feature Learning Using Co-Tuplet Loss for Offline Handwritten Signature Verification},
author = {Fu-Hsien Huang and Hsin-Min Lu},
year = {2023},
eprint = {2308.00428},
archivePrefix = {arXiv},
primaryClass = {cs.CV}
}