/Review-Handwritten-Signature-Verification

Review the state of art works of handwritten signature verification works

Review-Handwritten-Signature-Verification

Review the state of art works of handwritten signature verification works.

Depending on the acquisition method, signature verification systems are divided into: online (dynamic) and offline (static) with Writer Dependent or not categories.

  • In online case, an acquisition device, such as a digitizing table, is used to acquire the user’s signature. The data is collected as a sequence over time, containing the position of the pen, and in some cases including additional information such as the pen inclination\velocity\angle\curvature\full acceleration\coordinates\pressure 倾角\速度\角度\曲率\全加速度\坐标\压力
  • In offline signature verification, the signature is acquired after the writing process is completed. In this case, the signature is represented as a digital image or picture
  • Also, the can be devided into writer independent and writer dependent classification cases
  • SVM (Support Vector Machines) model is one of the most effective classifier for signature verification
  • Texture features (especially Local Binary Patterns LBP) is the best hand-crafted feature extractors in signature verification

Open source works

In many cases, the open source projects were deployed by matlab or python.

Offline project- Luizgh

Hafemann, L. G., Sabourin, R., & Oliveira, L. S. (2017, November).
[1] Offline handwritten signature verification—literature review
    URL: https://arxiv.org/pdf/1507.07909.pdf
[2] Learning Features for Offline Handwritten Signature Verification using Deep Convolutional Neural Networks
    URL: http://dx.doi.org/10.1016/j.patcog.2017.05.012 (preprint)
[3] Fixed-sized representation learning from Offline Handwritten Signatures of different sizes 
    URL: https://doi.org/10.1007/s10032-018-0301-6 (preprint)
  • Luizgh Offline Expert!
  • Project site: https://www.etsmtl.ca/Unites-de-recherche/LIVIA/Recherche-et-innovation/Projets/Signature-Verification
  • DCNN Deep convolutional neural network
  • SigNet adopted as the feature approach
    • wget "https://storage.googleapis.com/luizgh-datasets/models/signet_models.zip"
      • 3 signet models: Refere to paper [2]
        • signet.pkl; signetf_lambda0.95.pkl; and signetf_lambda0.999.pkl
      • Input scaned signatures need to be preprocess_signature() before the CNN. The aim is to make sure the input image/signature under the same resolution by wrap/crop/resize functions
    • wget "https://storage.googleapis.com/luizgh-datasets/models/signet_spp_models.zip"
      • 4 signet models: Refere to paper [3]
        • signet_spp_300dpi.pkl; signet_spp_300dpi_f.pkl and 600_dpi
        • (with or without f ) Training the model with or without forgeries
        • 300 or 600 is the scaned resolution of input image
        • 300dpi: input_size: (428, 612) and image_size: (476, 680)
        • 600dpi: input_size: (856, 1224) and image_size: (952, 1360)
      • Improve the preprocess_signature() by the introduced SPP methodology.
        • SPP the final layer in CNN, make sure CNN can handle any scale/size input image and generate the same sized feature vectors
  • Be aware:
    • Refer interactive_example.ipynb for SPP method
    • the input signature image picture is 1 color channel file gray color rather than RGB.
    • The time consumption during compile is high due to the limitation of Theano of DNN platform
  • *****questions need to be confirmed *****
    • Check the Euclidean Distance threshold in the paper (interactive_example.ipynb)
      • For actual classification, the simplest approach is computing the distance (in feature space) between a query signature and a reference. This is easy to implement (see the jupyter notebook example), but not very powerful.
      • Most commonly, people train Writer-dependent classifiers (one binary classifier for each user).

Offline projects

Online projects

Interesting Paper with identical name:

Online Signature Verification on Mobile Devices

Database Download-Able