Publications

Publications

Enhancing PLANETBRAIN every day

Enhancing PLANETBRAIN every day

Recent progress in the areas of Artificial Intelligence (AI) and Machine Learning (ML) are tremendous. Almost monthly, we see reports announcing breakthroughs in different technological aspects of AI.

As an organization focussing on research and development, we can look back on an increasing number of publications.

PUBLICATIONS

With its standardized MRI datasets of the entire spine, the German National Cohort (GNC) has the potential to deliver standardized biometric reference values for intervertebral discs (VD), vertebral bodies (VB) and spinal canal (SC). To handle such large-scale big data, artificial intelligence (AI) tools are needed. In this manuscript, we will present an AI software tool to analyze spine MRI and generate normative standard values. 330 representative GNC MRI datasets were randomly selected in equal distribution regarding parameters of age, sex and height. By using a 3D U-Net, an AI algorithm was trained, validated and tested. Finally, the machine learning algorithm explored the full dataset (n = 10,215). VB, VD and SC were successfully segmented and analyzed by using an AI-based algorithm. A software tool was developed to analyze spine-MRI and provide age, sex, and height-matched comparative biometric data. Using an AI algorithm, the reliable segmentation of MRI datasets of the entire spine from the GNC was possible and achieved an excellent agreement with manually segmented datasets. With the analysis of the total GNC MRI dataset with almost 30,000 subjects, it will be possible to generate real normative standard values in the future.

Authors: Felix Streckenbach (University Medical Center Rostock), Gundram Leifert (PLANET AI GmbH), Thomas Beyer (University Medical Center Rostock) et. al.

Journal: Healthcare 2022 (MDPI)

DOI: https://doi.org/10.3390/healthcare10112132

In contrast to Connectionist Temporal Classification (CTC) approaches, Sequence-To-Sequence (S2S) models for Handwritten Text Recognition (HTR) suffer from errors such as skipped or repeated words which often occur at the end of a sequence. In this paper, to combine the best of both approaches, we propose to use the CTC-Prefix-Score during S2S decoding. Hereby, during beam search, paths that are invalid according to the CTC confidence matrix are penalised. Our network architecture is composed of a Convolutional Neural Network (CNN) as visual backbone, bidirectional Long-Short-Term-Memory-Cells (LSTMs) as encoder, and a decoder which is a Transformer with inserted mutual attention layers. The CTC confidences are computed on the encoder while the Transformer is only used for character-wise S2S decoding. We evaluate this setup on three HTR data sets: IAM, Rimes, and StAZH. On IAM, we achieve a competitive Character Error Rate (CER) of 2.95% when pretraining our model on synthetic data and including a character-based language model for contemporary English. Compared to other state-of-the-art approaches, our model requires about 10–20 times less parameters. Access our shared implementations via this link to GitHub.

Authors: Christoph Wick (PLANET AI GmbH), Jochen Zöllner (PLANET AI GmbH, University of Rostock), Tobias Grüning (PLANET AI GmbH)

Series: DAS 2022 – 15th IAPR International Workshop on Document Analysis Systems

DOI: 10.1007/978-3-031-06555-2_18

Currently, the most widespread neural network architecture for training language models is the so-called BERT, which led to improvements in various Natural Language Processing (NLP) tasks. In general, the larger the number of parameters in a BERT model, the better the results obtained in these NLP tasks. Unfortunately, the memory consumption and the training duration drastically increases with the size of these models. In this article, we investigate various training techniques of smaller BERT models: We combine different methods from other BERT variants, such as ALBERT, RoBERTa, and relative positional encoding. In addition, we propose two new fine-tuning modifications leading to better performance: Class-Start-End tagging and a modified form of Linear Chain Conditional Random Fields. Furthermore, we introduce Whole-Word Attention, which reduces BERTs memory usage and leads to a small increase in performance compared to classical Multi-Head-Attention. We evaluate these techniques on five public German Named Entity Recognition (NER) tasks, of which two are introduced by this article.

Authors: Jochen Zöllner (PLANET AI GmbH, University of Rostock), Konrad Sperfeld (University of Rostock), Christoph Wick (PLANET AI GmbH), Roger Labahn (University of Rostock)

Journal: MDPI Information

DOI: 10.3390/info12110443

In order to apply Optical Character Recognition (OCR) to historical printings of Latin script fully automatically, we report on our efforts to construct a widely-applicable polyfont recognition model yielding text with a Character Error Rate (CER) around 2% when applied out-of-the-box. Moreover, we show how this model can be further finetuned to specific classes of printings with little manual and computational effort. The mixed or polyfont model is trained on a wide variety of materials, in terms of age (from the 15th to the 19th century), typography (various types of Fraktur and Antiqua), and languages (among others, German, Latin, and French). To optimize the results we combined established techniques of OCR training like pretraining, data augmentation, and voting. In addition, we used various preprocessing methods to enrich the training data and obtain more robust models. We also implemented a two-stage approach which first trains on all available, considerably unbalanced data and then refines the output by training on a selected more balanced subset. Evaluations on 29 previously unseen books resulted in a CER of 1.73%, outperforming a widely used standard model with a CER of 2.84% by almost 40%. Training a more specialized model for some unseen Early Modern Latin books starting from our mixed model led to a CER of 1.47%, an improvement of up to 50% compared to training from scratch and up to 30% compared to training from the aforementioned standard model. Our new mixed model is made openly available to the community.

Authors: Christian Reul (University of Würzburg), Christoph Wick (PLANET AI GmbH), Maximilian Nöth, Andreas Büttner, Maximilian Wehner (all University of Würzburg), Uwe Springmann (LMU München)

Series: ICDAR 2021

Pages: 112 – 126

DOI: 10.1007/978-3-030-86334-0_8

Most recently, Transformers – which are recurrent-free neural network architectures – achieved tremendous performances on various Natural Language Processing (NLP) tasks. Since Transformers represent a traditional Sequence-To-Sequence (S2S)-approach they can be used for several different tasks such as Handwritten Text Recognition (HTR). In this paper, we propose a bidirectional Transformer architecture for line-based HTR that is composed of a Convolutional Neural Network (CNN) for feature extraction and a Transformer-based encoder/decoder, whereby the decoding is performed in reading-order direction and reversed. A voter combines the two predicted sequences to obtain a single result. Our network performed worse compared to a traditional Connectionist Temporal Classification (CTC) approach on the IAM-dataset but reduced the state-of-the-art of Transformers-based approaches by about 25% without using additional data. On a signi cantly larger dataset, the proposed Transformer significantly outperformed our reference model by about 26%. In an error analysis, we show that the Transformer is able to learn a strong language model which explains why a larger training dataset is required to outperform traditional approaches and discuss why Transformers should be used with caution for HTR due to several shortcomings such as repetitions in the text.

Authors: Christoph Wick (PLANET AI GmbH), Jochen Zöllner (PLANET AI GmbH, University of Rostock), Tobias Grüning (PLANET AI GmbH)

Series: ICDAR 2021

Pages: 112 – 126

In this paper, we propose a novel method for Automatic Text Recognition (ATR) on early printed books. Our approach significantly reduces the Character Error Rates (CERs) for book-specific training when only a few lines of Ground Truth (GT) are available and considerably outperforms previous methods. An ensemble of models is trained simultaneously by optimising each one independently but also with respect to a fused output obtained by averaging the individual confidence matrices. Various experiments on five early printed books show that this approach already outperforms the current state-of-the-art by up to 20% and 10% on average. Replacing the averaging of the confidence matrices during prediction with a con dence-based voting boosts our results by an additional 8% leading to a total average improvement of about 17%.

Authors: Christoph Wick (PLANET AI GmbH), Christian Reul (University of Würzburg)

Series: ICDAR 2021

Pages: 385 – 399

DOI: 10.1007/978-3-030-86549-8_25

tfaip is a Python-based research framework for developing, structuring, and deploying DeepLearning projects powered by Tensorflow (Abadi et al., 2015) and is intended for scientists of universities or organizations who research, develop, and optionally deploy Deep Learning models. tfaip enables both simple and complex implementation scenarios, such as image classification, object detection, text recognition, natural language processing, or speech recognition. Each scenario is highly configurable by parameters that can directly be modified by the command line or the API.

Authors: Christoph Wick, Benjamin Kühn, Gundram Leifert (all PLANET AI GmbH), Konrad Sperfeld (CITlab, University of Rostock), Jochen Zöllner (PLANET AI GmbH, University of Rostock), Tobias Grüning (PLANET AI GmbH)

Journal: The Journal of Open Source Software (JOSS)

DOI: 10.21105/joss.03297

Automated text recognition is a fundamental problem in Document Image Analysis. Optical models are used for modeling characters while language models are used for composing sentences. Since the scripts and linguistic context differ widely, it is mandatory to specialize the models by training on task-dependent ground-truth. However, to create a sufficient amount of ground-truth, at least for historical handwritten scripts, well-qualified persons have to mark and transcribe text lines, which is very time-consuming. On the other hand, in many cases unassigned transcripts are already available on page-level from another process chain, or at least transcripts from similar linguistic context are available. In this work we present two approaches that make use of such transcripts: whereas the first one creates training data by automatically assigning page-dependent transcripts to text lines, the second one uses a task-specific language model to generate highly confident training data. Both approaches are successfully applied on a very challenging historical handwritten collection.

Authors: Gundram Leifert (PLANET AI GmbH), Joan Andreu Sànchez (Pattern Recognition and Human Language Technologies Center), Roger Labahn (Computational Intelligence Technology Lab)

Series: ICFHR ’20

Pages: To appear

Note: This work was partially funded by the Generalitat Valenciana under the EU-FEDER Comunitat Valenciana 2014-2020 grant IDIFEDER/2018/025 “Sistemas de fabricación inteligente para la indústria 4.0”. | in proceeding

Encoder-decoder models have become an effective approach for sequence learning tasks like machine translation, image captioning and speech recognition, but have yet to show competitive results for handwritten text recognition. To this end, we propose an attention-based sequence-to-sequence model. It combines a convolutional neural network as a generic feature extractor with a recurrent neural network to encode both the visual information, as well as the temporal context between characters in the input image, and uses a separate recurrent neural network to decode the actual character sequence. We make experimental comparisons between various attention mechanisms and positional encodings, in order to find an appropriate alignment between the input and output sequence. The model can be trained end-to-end and the optional integration of a hybrid loss allows the encoder to retain an interpretable and usable output, if desired. We achieve competitive results on the IAM and ICFHR2016 READ data sets compared to the state-of-the-art without the use of a language model, and we significantly improve over any recent sequence-to-sequence approaches.

Authors: Johannes Michael, Roger Labahn, Tobias Grüning, Jochen Zöllner

Booktitle: Proceedings of the 2019 15th International Conference on Document Analysis and Recognition

Series: ICDAR ’19

Pages: To appear

Note: Partially funded by the European Unions Horizon 2020 research and innovation programme under grant agreement No 674943 (READ) | in proceeding

Measuring the performance of text recognition and text line detection engines is an important step to objectively compare systems and their configuration. There exist well-established measures for both tasks separately. However, there is no sophisticated evaluation scheme to measure the quality of a combined text line detection and text recognition system. The F-measure on word level is a well-known methodology, which is sometimes used in this context. Nevertheless, it does not take into account the alignment of hypothesis and ground truth text and can lead to deceptive results. Since users of automatic information retrieval pipelines in the context of text recognition are mainly interested in the end-to-end performance of a given system, there is a strong need for such a measure. Hence, we present a measure to evaluate the quality of an end-to-end text recognition system. The basis for this measure is the well established and widely used character error rate, which is limited — in its original form — to aligned hypothesis and ground truth texts. The proposed measure is flexible in a way that it can be configured to penalize different reading orders between the hypothesis and ground truth and can take into account the geometric position of the text lines. Additionally, it can ignore over- and under- segmentation of text lines. With these parameters it is possible to get a measure fitting best to its own needs.

Authors: Gundram Leifert, Roger Labahn, Tobias Grüning, Svenja Leifert

Booktitle: Proceedings of the 2019 15th International Conference on Document Analysis and Recognition

Series: ICDAR ’19

Pages: To appear

Note: Partially funded by the European Unions Horizon 2020 research and innovation programme under grant agreement No 674943 (READ) | in proceeding

We present a recognition and retrieval system for the ICDAR2017 Competition on Information Extraction in Historical Handwritten Records which successfully infers person names and other data from marriage records. The system extracts information from the line images with a high accuracy and outperforms the baseline. The optical model is based on Neural Networks. To infer the desired information, regular expressions are used to describe the set of feasible words sequences.

Authors: Tobias Strauß, Max Weidemann, Johannes Michael, Gundram Leifert, Tobias Grüning, Roger Labahn

Journal: CoRR

Volume: abs/1804.09943

Note: Partially funded by the European Unions Horizon 2020 research and innovation programme under grant agreement No 674943 (READ)

Accessibility of the valuable cultural heritage which is hidden in countless scanned historical documents is the motivation for the presented dissertation. The developed (fully automatic) text line extraction methodology combines state-of-the-art machine learning techniques and modern image processing methods. It demonstrates its quality by outperforming several other approaches on a couple of benchmarking datasets. The method is already being used by a wide audience of researchers from different disciplines and thus contributes its (small) part to the aforementioned goal.

Author: Tobias Grüning

Type: PhD thesis

School: Universität Rostock

Author: Tobias Grüning, Roger Labahn, Markus Diem, Florian Kleber, Stefan Fiel

Booktitle: 2018 13th IAPR International Workshop on Document Analysis Systems (DAS)

Pages: 351-356

Note: Partially funded by the European Unions Horizon 2020 research and innovation programme under grant agreement No 674943 (READ) | in proceeding

DOI: 10.1109/DAS.2018.38

Authors: Tobias Grüning, Gundram Leifert, Tobias Strauß, Roger Labahn

Booktitle: 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR)

Volume: 01

Pages: 351-356

Note: Partially funded by the European Unions Horizon 2020 research and innovation programme under grant agreement No 674943 (READ) | inproceeding

DOI: 10.1109/ICDAR.2017.47

In der Handschrifterkennung geben neuronale Netze Folgen von Wahrscheinlichkeit pro Character aus. Gegenstand der Arbeit ist das Optimierungsproblem, die Ausgaben neuronaler Netzwerke in Maschinen-lesbare Texte zu konvertieren. Dies wird mit Hilfe von gewichteten Automaten realisiert. Als wesentliches Resultat wird eine effiziente Heuristik entwickelt, die die wahrscheinlichste Buchstabenfolge aller durch reguläre Ausdrücke beschränkter Folgen findet.

Author: Tobias Strauß

Type: PhD thesis

School: Universität Rostock

Authors: Tobias Grüning, Gundram Leifert, Tobias Strauß, Roger Labahn

Booktitle: CLEF2016 Working Notes

Series: CEUR Workshop Proceedings

Publisher: CEUR-WS.org

Pages: 351-356

Note: Partially funded by grant no. KF2622304SS3 (Kooperationsprojekt) in Zentrales Innovationsprogramm Mittelstand (ZIM) by Bundesrepublik Deutschland (BMWi) and the European Unions Horizon 2020 research and innovation programme under grant agreement No 674943 (READ) | in proceeding

DOI: 10.1109/ICDAR.2017.47

The transcription of handwritten text on images is one task in machine learning and one solution to solve it is using multi- dimensional recurrent neural networks (MDRNN) with connectionist temporal classification (CTC). The RNNs can contain special units, the long short-term memory (LSTM) cells. They are able to learn long term dependencies but they get unstable when the dimension is chosen greater than one. We defined some useful and necessary properties for the one-dimensional LSTM cell and extend them in the multi-dimensional case. Thereby we introduce several new cells with better stability. We present a method to design cells using the theory of linear shift invariant systems. The new cells are compared to the LSTM cell on the IFN/ENIT and Rimes database, where we can improve the recognition rate compared to the LSTM cell. So each application where the LSTM cells in MDRNNs are used could be improved by substituting them by the new developed cells.

Authors: Gundram Leifert, Tobias Strauß, Tobias Grüning, Welf Wustlich, Roger Labahn

Journal: Journal of Machine Learning Research

Volume: 17

Number: 97

Pages: 1-37

This article proposes a convenient tool for decoding the output of neural networks trained by Connectionist Temporal Classification (CTC) for handwritten text recognition. We use regular expressions to describe the complex structures expected in the writing. The corresponding finite automata are employed to build a decoder. We analyze theoretically which calculations are relevant and which can be avoided. A great speed-up results from an approximation. We conclude that the approximation most likely fails if the regular expression does not match the ground truth which is not harmful for many applications since the low probability will be even underestimated. The proposed decoder is very efficient compared to other decoding methods. The variety of applications reaches from information retrieval to full text recognition. We refer to applications where we integrated the proposed decoder successfully.

Authors: Tobias Strauß, Gundram Leifert, Tobias Grüning, Roger Labahn

Journal: Neural Networks

Volume: 79

Pages: 1 – 11

Note: Partially funded by grant no. KF2622304SS3 (Kooperationsprojekt) in Zentrales Innovationsprogramm Mittelstand (ZIM) by Bundesrepublik Deutschland (BMWi)

We describe CITlab’s recognition system for the HTRtS competition attached to the 13. International Conference on Document Analysis and Recognition, ICDAR 2015. The task comprises the recognition of historical handwritten documents. The core algorithms of our system are based on multi-dimensional recurrent neural networks (MDRNN) and connectionist temporal classification (CTC). The software modules behind that as well as the basic utility technologies are essentially powered by PLANET’s ARGUS framework for intelligent text recognition and image processing.

Authors: Gundram Leifert, Tobias Strauß, Tobias Grüning, Roger Labahn

Journal: CoRR

Volume: abs/1605.08412

Note: Partially funded by the European Unions Horizon 2020 research and innovation programme under grant agreement No 674943 (READ)

We describe CITlab’s recognition system for the HTRtS competition attached to the 14. International Conference on Frontiers in Handwriting Recognition, ICFHR 2014. The task comprises the recognition of historical handwritten documents. The core algorithms of our system are based on multi-dimensional recurrent neural networks (MDRNN) and connectionist temporal classification (CTC). The software modules behind that as well as the basic utility technologies are essentially powered by PLANET’s ARGUS framework for intelligent text recognition and image processing.

Authors: Tobias Strauß, Tobias Grüning, Gundram Leifert, Roger Labahn

Journal: CoRR

Volume: abs/1412.3949

Note: Partially funded by research grant no. V220-630-08-TFMV-S/F-059 (Verbundvorhaben, Technologieförderung Land Mecklenburg-Vorpommern) in European Social / Regional Development Funds

We describe CITlab’s recognition system for the ANWRESH-2014 competition attached to the 14. International Conference on Frontiers in Handwriting Recognition, ICFHR 2014. The task comprises word recognition from segmented historical documents. The core components of our system are based on multi-dimensional recurrent neural networks (MDRNN) and connectionist temporal classification (CTC). The software modules behind that as well as the basic utility technologies are essentially powered by PLANET’s ARGUS framework for intelligent text recognition and image processing.

Authors: Tobias Strauß, Tobias Grüning, Gundram Leifert, Roger Labahn

Journal: CoRR

Volume: abs/1412.6012

Note: Partially funded by research grant no. V220-630-08-TFMV-S/F-059 (Verbundvorhaben, Technologieförderung Land Mecklenburg-Vorpommern) in European Social / Regional Development Funds

In the recent years it turned out that multidimensional recurrent neural networks (MDRNN) perform very well for offline handwriting recognition tasks like the OpenHaRT 2013 evaluation DIR. With suitable writing preprocessing and dictionary lookup, our ARGUS software completed this task with an error rate of 26.27% in its primary setup.

Authors: Tobias Strauß, Tobias Grüning, Gundram Leifert, Roger Labahn

Journal: CoRR

Volume: abs/1412.6061

Note: Partially funded by research grant no. V220-630-08-TFMV-S/F-059 (Verbundvorhaben, Technologieförderung Land Mecklenburg-Vorpommern) in European Social / Regional Development Funds

This article develops approaches to generate dynamical reservoirs of echo state networks with desired properties reducing the amount of randomness. It is possible to create weight matrices with a predefined singular value spectrum. The procedure guarantees stability (echo state property). We prove the minimization of the impact of noise on the training process. The resulting reservoir types are strongly related to reservoirs already known in the literature. Our experiments show that well-chosen input weights can improve performance.

Authors: Tobias Strauß, Welf Wustlich, Roger Labahn

Journal: Neural Computation

Volume: 24

Number: 12

Pages: 3246-3276

Note: Partially funded by the research grant no. V220-630-08-TFMV-S/F-059 (Verbundvorhaben, Technologieförderung Land Mecklenburg-Vorpommern) in European Social / Regional Development Funds