Attention based Handwriting Verification

Conference paper presented in ICFHR2020, Dortmund, Germany

https://arxiv.org/abs/2009.04532

Prediction Example

More Results

Correspondence region maps from Cross Attention

1. Images from same writer where the model is able to locate some points corresponding to query
2. Images from different writers where the model is unable to locate any point similar to queried pixel region

Salient region maps from Soft Attention

1. Salient regions when the model assumes that the writers are same. Model highlights the similar looking regions.
2. Salient regions when the model assumes that the writers are different. Model highlights the dissimilar looking regions.

Citation

@INPROCEEDINGS{
  9257763
  ,  author={M. A. {Shaikh} and T. {Duan} and M. {Chauhan} and S. N. {Srihari}}
  ,  booktitle={2020 17th International Conference on Frontiers in Handwriting Recognition (ICFHR)}
  ,  title={Attention based Writer Independent Verification}
  ,  year={2020}
  ,  pages={373-379}
  ,  doi={10.1109/ICFHR2020.2020.00074}
}

Datasets

• Handwritten "AND": https://github.com/mshaikh2/HDL_Forensics
• CEDAR Signature: http://www.cedar.buffalo.edu/NIJ/data