The Computer Vision Center and the Qidenus Group Company open two industrial PhD positions on pattern recognition and computer vision techniques applied to information extraction from large scale historical document image sources.
The context: information extraction from EU demographic document sources
Historical documents reflect the identities and contexts of the past, and the access to their contents allows citizens at large to know their individual and collective memory. Search centered at people is very important in historical research, including family history and genealogical research. Genealogical documents such as birth, marriage or dead certificates or census records contain information that reflect a picture of a historical context: a person’s life, an event, a location at some period of time. The CVC and Qidenus are currently working in the development of computational methods for the massive analysis of digitized historical demographic documents.
Research environment: academia and industry
Successful candidates will become doctoral students at the Universitat Autònoma de Barcelona, where the CVC is located. The students will be therefore based in Barcelona, with regular visits to the Qidenus headquarters in Vienna or Berlin. The students will join the group of Document Analysis of the CVC, consisting of 20 researchers. In addition, a supervision from the industrial side will be provided.
(1 position per topic)
Semantic annotation of handwritten word images. Inspired in approaches like [1,2,3] that learn a joint embedding space that associates visual features extracted from images to concepts. We plan to use deep neural networks to associate semantic categories to named entities (family names, dates, occupations, etc.) .
Multiwriter word recognition and spotting using transfer learning models. There are some words that appear in different documents written by different writers. The traditional HTR techniques use to decrease the performance if the writing style of the recognition set is not the same as the training set. We propose to investigate the models of writer adaptation and transfer learning proposed in works like [5,6].
 Albert Gordo, Jon Almazan, Naila Murray, Florent Perronnin. LEWIS: Latent Embeddings for Word Images and their Semantics. ICCV’2015.
 Arik Poznanski, Lior Wolf. CNN-N-Gram for Handwriting Word Recognition. CVPR’2016
 Sebastian Sudholt, Gernot A. Fink. PHOCNet: A Deep Convolutional Neural Network for Word Spotting in Handwritten Documents. arXiv2016. Submitted to ICFHR’2016.
 J.Ignacio Toledo, Sebastian Sudholt, Alicia Fornés, Jordi Cucurull, Gernot A. Fink, and Josep Lladós. Handwritten Word Image Categorization with Convolutional Neural Networks and Spatial Pyramid Pooling. Submitted to S+SSPR’2016.
 Xu-Yao Zhang, Cheng-Lin Liu. Style transfer matrix learning for writer adaptation. CVPR’2011.
 X.-Y. Zhang, C.-L. Liu, Writer adaptation with style transfer mapping, IEEE Trans. Pattern Analysis and Machine Intelligence, 35(7): 1773-1787, 2013.
The scholarship is given for 3 years provided successful progress and should lead to a final PhD dissertation. The candidate must perform high quality research leading to the publication of papers in well-known international conferences and journals with high impact factor. Since it is an industrial PhD, the candidate will transfer the results of the research to the application framework of the sponsoring company.
Annual gross remuneration: €22,000 (additionally the academic expenses, and travel costs between the laboratory and the company will be covered).
Qualifications and skills required
- Bachelor degree in Computer Science or a related field such as electrical, telecommunications engineering, Mathematics or Physics.
- Master degree in Computer Vision, Pattern Recognition or closely related topics.
- Good mathematical understanding.
- High motivation for research.
- Capability of working in an autonomous way.
- Good programming skills (preferably C++, Python, Matlab).
- Good communication skills in English, both in written and oral form.
- Machine Learning, Pattern Recognition and Document Analysis research experience will be an asset.
How to apply:
For the first topic, Semantic annotation of handwritten word images:
For the second topic, Multiwriter word recognition and spotting using transfer learning models:
Send required information to:
Campus UAB, Edifici O