As for many applications of computer vision, significant progress have been achieved recently in the field of handwriting recognition, thanks to Deep Learning. Handwriting recognition systems are very similar to speech recognition ones: convolutional neural networks are used to extract features, recurrent neural networks model the sequences of characters and attention mechanisms are used to follow the writing and decode the sequence of words. Similarly to other applications of deep learning, the performance are strongly correlated to the quantity of good quality annotated data. However, when working with historical documents, collecting a large quantity of annotated samples from handwritten text may be difficult: transcribing documents in latin or old languages can only be done by paleographers. New training approaching, taking advantage of partially transcribed documents or being able to adapt in an unsupervised ways to a homogeneous collection of document must be developed.

Missions

The workplan is the following :

  • First, a survey of the different Deep Learning architectures recently proposed for handwriting recognition will be conducted.

  • In collaboration with other Machine Learning engineers, a handwriting recognition system will be implemented and compared to state-of-the-art on standard databases (RIMES, IAM , READ [Arora2019, Puigcerver2017, Strauß2018]).

  • Self-supervised [Trinh2019] and self-training [Kahn2019] method will be studied, implemented and evaluated in the framework of several current historical research projects conducted at TEKLIA (ANR HORAE, JPI HOME, Balsac)

Environment

  • Python, pytorch/keras, kaldi, linux, GPU-servers.
  • Regular reading groups are organized to discuss research papers.
  • International collaboration with European research partners.

Place

TEKLIA, 30 rue Raymond Losserand, 75015 Paris

Contact

Christopher Kermorvant : kermorvant@teklia.com

Bibliography

  • [Arora2019] A. Arora et al., “Using ASR methods for OCR,” in International Conference of Document Analysis and Recognition, 2019.
  • [Bluche2017] T. Bluche, C. Kermorvant, and H. Ney, How to design deep neural networks for handwriting recognition. 2017.
  • [Kahn2019] J. Kahn, A. Lee, and A. Hannun, "Self-training for end-to-end speech recognition", 2019.
  • [Puigcerver2017] J. Puigcerver, “Are Multidimensional Recurrent Layers Really Necessary for Handwritten Text Recognition?,” 2017 14th IAPR Int. Conf. Doc. Anal. Recognit., pp. 67–72, 2017.
  • [Strauß2018] T. Strauß, G. Leifert, R. Labahn, and G. Mühlberger, “ICFHR2018 Competition on Automated Text Recognition on a READ Dataset,” in International Conference on Frontiers in Handwriting Recognition, 2018.
  • [Trinh2019] T. H. Trinh, M.-T. Luong, and Q. V. Le, “Selfie: Self-supervised Pretraining for Image Embedding” 2019.