Refine
Has Fulltext
- yes (21)
Is part of the Bibliography
- yes (21)
Year of publication
Document Type
- Doctoral Thesis (19)
- Master Thesis (2)
Keywords
- Data Mining (3)
- Maschinelles Lernen (3)
- Deep Learning (2)
- Information Extraction (2)
- Optical Character Recognition (2)
- ALS (1)
- Alter Druck (1)
- Anomalieerkennung (1)
- Arbeitsablaufplanung (1)
- Automatic Text Reconition (1)
In dieser Arbeit wurde ein Kollektiv chronisch herzinsuffizienter Patienten aus der niedergelassenen kardiologischen Betreuung in Bayern analysiert und auf die Umsetzung der zum Zeitpunkt der HF-Bavaria Studie gültigen Leitlinien untersucht. Dabei wurde das Patientenkollektiv nach dem Geschlecht und zusätzlich auch nach den neu definierten Herzinsuffizienz-Klassen der aktuell gültigen Leitlinien eingeteilt, um Unterschiede und Gemeinsamkeiten innerhalb dieser Differenzierungen darstellen zu können und einen Vergleich zu den Studien der jüngeren Vergangenheit zu ermöglichen.
Die Patienten der HF-Bavaria Studie waren zu 65,9 % männlich (n = 3569) und zu 34,1 % weiblich (n = 1848). Die Frauen litten häufiger unter HFpEF, waren seit kürzerer Zeit herzinsuffizient und waren in der Vergangenheit seltener zur Therapieintensivierung oder Intervention hospitalisiert. Die Patientinnen berichteten dabei weniger häufig Komorbiditäten. So fanden sich bei den Frauen seltener KHK, Niereninsuffizienz oder Diabetes mellitus, hingegen häufiger Herzklappenerkrankungen und Vorhofflimmern. Weiterhin wurden die Patientinnen weniger häufig mit ACE-Hemmer, Betablocker und MRA, dagegen häufiger mit ARB und Digitalis behandelt.
Im Patientenkollektiv der HF-Bavaria Studie hatten 29,0 % eine HFrEF (n = 1581), 28,9 % eine HFmrEF (n = 1577) und 42,0 % eine HFpEF (n = 2291). Patienten mit HFrEF waren überwiegend männlich, zum größten Teil seit mehr als 5 Jahren herzinsuffizient und im Vergleich zu den anderen Herzinsuffizienz-Klassen häufiger in den NYHA-Stadien III und IV eingestuft. HFrEF Patienten hatten den größten Anteil an bereits erfolgten Interventionen und Device-Therapien und die durchschnittlich höchste Anzahl an Komorbiditäten. Das Komorbiditätenspektrum bei Patienten mit HFmrEF lag prozentual in den meisten Kategorien zwischen den beiden anderen Herzinsuffizienz-Klassen. Patienten mit HFpEF waren überThe ewiegend weiblich, wiesen vergleichsweise am häufigsten eine komorbide Hypertonie oder ein Vorhofflimmern auf, während eine KHK deutlich seltener vorlag, als es in den anderen Herzinsuffizienz-Klassen der Fall war.
Die Prüfung der leitliniengerechten Pharmakotherapie bei HFrEF-Patienten ergab eine insgesamt gleichwertige Verschreibungshäufigkeit im geschlechtsspezifischen Vergleich der nach NYHA-Stadium indizierten Medikamentenklassen und Kombinationstherapien. Lediglich im NYHA-Stadium III konnte gezeigt werden, dass Männer signifikant häufiger mit einem Betablocker therapiert wurden. Weiterhin zeigte sich, bis auf wenige Ausnahmen, eine auch im nationalen und internationalen Vergleich hohe prozentuale Einnahme der stadienabhängig indizierten Medikamente. Die Einnahmerate von MRAs war vergleichsweise noch geringer als zu erwarten wäre, jedoch konnte das begleitende Vorliegen relevanter Kontraindikationen nicht zuverlässig genug erfasst werden, um die tatsächliche Versorgungslücke zu quantifizieren.
Die Analyse der Pharmakotherapie von HFmrEF- und HFpEF-Patienten zeigte, trotz bisher fehlender wissenschaftlicher Erkenntnisse zur optimalen medikamentösen Therapie dieser Patientengruppen, sehr ähnliche Einnahmehäufigkeiten der verschiedenen Substanzklassen im Vergleich zu den HFrEF-Patienten.
Die Therapie mit Devices war im Patientenkollektiv der HF-Bavaria Studie vergleichsweise selten und dabei häufiger bei männlichen Patienten vorzufinden. Die Analyse der leitliniengetreuen Indikationen von ICDs, CRTs und CRT-ICDs zu den tatsächlich implantierten Devices ergab Hinweise auf eine Unterversorgung vermittels apparativer Therapiemöglichkeiten.
Die Auswertung der HF-Bavaria Studie bestätigte die von uns erwartete Heterogenität und Komplexität der herzinsuffizienten Patienten in der niedergelassenen kardiologischen Betreuung. In dieser Arbeit konnte gezeigt werden, dass bedeutsame Unterschiede im Hinblick auf das Profil, den Verlauf und die Therapie von männlichen und weiblichen herzinsuffizienten Patienten bestehen. Die Therapieempfehlungen der Leitlinien richten sich trotz dieser Unterschiede vorrangig nach der Herzinsuffizienz-Klasse der Patienten. Bisher existierten in den Leitlinien vorrangig Therapieempfehlungen für Patienten mit einer HFrEF (und LVEF <40 %). Im Patientenkollektiv fanden sich jedoch zu 71 % Patienten mit einer LVEF ≥40 %. Dies bedeutet, dass für den Großteil der Patienten in unserer Studie bisher keine evidenzbasierten Behandlungsalgorithmen existieren, insbesondere zur Pharmakotherapie. Künftig sollte die Forschung vermehrt auf diese Evidenzlücken eingehen und idealerweise eine personalisierte Therapie ermöglichen.
Abschließend lässt sich feststellen, dass die leitliniengerechte Therapie der herzinsuffizienten Patienten in der niedergelassenen kardiologischen Versorgung in Bayern eine im nationalen und internationalen Kontext fortgeschrittene Qualität besitzt. Dennoch wurden erwartungsgemäß Möglichkeiten zur Qualitätsverbesserung im vorliegenden Projekt identifiziert.
Deep learning enables enormous progress in many computer vision-related tasks. Artificial Intel- ligence (AI) steadily yields new state-of-the-art results in the field of detection and classification. Thereby AI performance equals or exceeds human performance. Those achievements impacted many domains, including medical applications.
One particular field of medical applications is gastroenterology. In gastroenterology, machine learning algorithms are used to assist examiners during interventions. One of the most critical concerns for gastroenterologists is the development of Colorectal Cancer (CRC), which is one of the leading causes of cancer-related deaths worldwide. Detecting polyps in screening colonoscopies is the essential procedure to prevent CRC. Thereby, the gastroenterologist uses an endoscope to screen the whole colon to find polyps during a colonoscopy. Polyps are mucosal growths that can vary in severity.
This thesis supports gastroenterologists in their examinations with automated detection and clas- sification systems for polyps. The main contribution is a real-time polyp detection system. This system is ready to be installed in any gastroenterology practice worldwide using open-source soft- ware. The system achieves state-of-the-art detection results and is currently evaluated in a clinical trial in four different centers in Germany.
The thesis presents two additional key contributions: One is a polyp detection system with ex- tended vision tested in an animal trial. Polyps often hide behind folds or in uninvestigated areas. Therefore, the polyp detection system with extended vision uses an endoscope assisted by two additional cameras to see behind those folds. If a polyp is detected, the endoscopist receives a vi- sual signal. While the detection system handles the additional two camera inputs, the endoscopist focuses on the main camera as usual.
The second one are two polyp classification models, one for the classification based on shape (Paris) and the other on surface and texture (NBI International Colorectal Endoscopic (NICE) classification). Both classifications help the endoscopist with the treatment of and the decisions about the detected polyp.
The key algorithms of the thesis achieve state-of-the-art performance. Outstandingly, the polyp detection system tested on a highly demanding video data set shows an F1 score of 90.25 % while working in real-time. The results exceed all real-time systems in the literature. Furthermore, the first preliminary results of the clinical trial of the polyp detection system suggest a high Adenoma Detection Rate (ADR). In the preliminary study, all polyps were detected by the polyp detection system, and the system achieved a high usability score of 96.3 (max 100). The Paris classification model achieved an F1 score of 89.35 % which is state-of-the-art. The NICE classification model achieved an F1 score of 81.13 %.
Furthermore, a large data set for polyp detection and classification was created during this thesis. Therefore a fast and robust annotation system called Fast Colonoscopy Annotation Tool (FastCAT) was developed. The system simplifies the annotation process for gastroenterologists. Thereby the
i
gastroenterologists only annotate key parts of the endoscopic video. Afterward, those video parts are pre-labeled by a polyp detection AI to speed up the process. After the AI has pre-labeled the frames, non-experts correct and finish the annotation. This annotation process is fast and ensures high quality. FastCAT reduces the overall workload of the gastroenterologist on average by a factor of 20 compared to an open-source state-of-art annotation tool.
Environmental issues have emerged especially since humans burned fossil fuels, which led to air pollution and climate change that harm the environment. These issues’ substantial consequences evoked strong efforts towards assessing the state of our environment.
Various environmental machine learning (ML) tasks aid these efforts. These tasks concern environmental data but are common ML tasks otherwise, i.e., datasets are split (training, validatition, test), hyperparameters are optimized on validation data, and test set metrics measure a model’s generalizability. This work focuses on the following environmental ML tasks: Regarding air pollution, land use regression (LUR) estimates air pollutant concentrations at locations where no measurements are available based on measured locations and each location’s land use (e.g., industry, streets). For LUR, this work uses data from London (modeled) and Zurich (measured). Concerning climate change, a common ML task is model output statistics (MOS), where a climate model’s output for a study area is altered to better fit Earth observations and provide more accurate climate data. This work uses the regional climate model (RCM) REMO and Earth observations from the E-OBS dataset for MOS. Another task regarding climate is grain size distribution interpolation where soil properties at locations without measurements are estimated based on the few measured locations. This can provide climate models with soil information, that is important for hydrology. For this task, data from Lower Franconia is used.
Such environmental ML tasks commonly have a number of properties: (i) geospatiality, i.e., their data refers to locations relative to the Earth’s surface. (ii) The environmental variables to estimate or predict are usually continuous. (iii) Data can be imbalanced due to relatively rare extreme events (e.g., extreme precipitation). (iv) Multiple related potential target variables can be available per location, since measurement devices often contain different sensors. (v) Labels are spatially often only sparsely available since conducting measurements at all locations of interest is usually infeasible. These properties present challenges but also opportunities when designing ML methods for such tasks.
In the past, environmental ML tasks have been tackled with conventional ML methods, such as linear regression or random forests (RFs). However, the field of ML has made tremendous leaps beyond these classic models through deep learning (DL). In DL, models use multiple layers of neurons, producing increasingly higher-level feature representations with growing layer depth. DL has made previously infeasible ML tasks feasible, improved the performance for many tasks in comparison to existing ML models significantly, and eliminated the need for manual feature engineering in some domains due to its ability to learn features from raw data. To harness these advantages for environmental domains it is promising to develop novel DL methods for environmental ML tasks.
This thesis presents methods for dealing with special challenges and exploiting opportunities inherent to environmental ML tasks in conjunction with DL. To this end, the proposed methods explore the following techniques: (i) Convolutions as in convolutional neural networks (CNNs) to exploit reoccurring spatial patterns in geospatial data. (ii) Posing the problems as regression tasks to estimate the continuous variables. (iii) Density-based weighting to improve estimation performance for rare and extreme events. (iv) Multi-task learning to make use of multiple related target variables. (v) Semi–supervised learning to cope with label sparsity. Using these techniques, this thesis considers four research questions: (i) Can air pollution be estimated without manual feature engineering? This is answered positively by the introduction of the CNN-based LUR model MapLUR as well as the off-the-shelf LUR solution OpenLUR. (ii) Can colocated pollution data improve spatial air pollution models? Multi-task learning for LUR is developed for this, showing potential for improvements with colocated data. (iii) Can DL models improve the quality of climate model outputs? The proposed DL climate MOS architecture ConvMOS demonstrates this. Additionally, semi-supervised training of multilayer perceptrons (MLPs) for grain size distribution interpolation is presented, which can provide improved input data. (iv) Can DL models be taught to better estimate climate extremes? To this end, density-based weighting for imbalanced regression (DenseLoss) is proposed and applied to the DL architecture ConvMOS, improving climate extremes estimation. These methods show how especially DL techniques can be developed for environmental ML tasks with their special characteristics in mind. This allows for better models than previously possible with conventional ML, leading to more accurate assessment and better understanding of the state of our environment.
Serverless computing is an emerging cloud computing paradigm that offers a highlevel
application programming model with utilization-based billing. It enables the
deployment of cloud applications without managing the underlying resources or
worrying about other operational aspects. Function-as-a-Service (FaaS) platforms
implement serverless computing by allowing developers to execute code on-demand
in response to events with continuous scaling while having to pay only for the
time used with sub-second metering. Cloud providers have further introduced
many fully managed services for databases, messaging buses, and storage that also
implement a serverless computing model. Applications composed of these fully
managed services and FaaS functions are quickly gaining popularity in both industry
and in academia.
However, due to this rapid adoption, much information surrounding serverless
computing is inconsistent and often outdated as the serverless paradigm evolves.
This makes the performance engineering of serverless applications and platforms
challenging, as there are many open questions, such as: What types of applications
is serverless computing well suited for, and what are its limitations? How should
serverless applications be designed, configured, and implemented? Which design
decisions impact the performance properties of serverless platforms and how can
they be optimized? These and many other open questions can be traced back to an
inconsistent understanding of serverless applications and platforms, which could
present a major roadblock in the adoption of serverless computing.
In this thesis, we address the lack of performance knowledge surrounding serverless
applications and platforms from multiple angles: we conduct empirical studies
to further the understanding of serverless applications and platforms, we introduce
automated optimization methods that simplify the operation of serverless applications,
and we enable the analysis of design tradeoffs of serverless platforms by
extending white-box performance modeling.
Die nächtliche (24-stündige) Überwachung des intraokularen Drucks (IOD) bei stationären Glaukompatienten wird in Europa seit mehr als 100 Jahren eingesetzt, um Spitzenwerte zu erkennen, die während der regulären Sprechstundenzeiten übersehen werden. Daten, die diese Praxis unterstützen, fehlen, zum Teil weil es schwierig ist, manuell erstellte IOD-Kurven mit dem objektiven Verlauf des Glaukoms zu korrelieren. Um dieses Problem zu beheben, haben wir automatisierte IOD-Datenextraktionswerkzeuge eingesetzt und auf eine Korrelation mit einem fortschreitenden Verlust der retinalen Nervenfaserschicht auf der optischen Kohärenztomographie im Spektralbereich (SDOCT) getestet.
Detecting anomalies in transaction data is an important task with a high potential to avoid financial loss due to irregularities deliberately or inadvertently carried out, such as credit card fraud, occupational fraud in companies or ordering and accounting errors. With ongoing digitization of our world, data-driven approaches, including machine learning, can draw benefit from data with less manual effort and feature engineering. A large variety of machine learning-based anomaly detection methods approach this by learning a precise model of normality from which anomalies can be distinguished. Modeling normality in transactional data, however, requires to capture distributions and dependencies within the data precisely with special attention to numerical dependencies such as quantities, prices or amounts.
To implicitly model numerical dependencies, Neural Arithmetic Logic Units have been proposed as neural architecture. In practice, however, these have stability and precision issues.
Therefore, we first develop an improved neural network architecture, iNALU, which is designed to better model numerical dependencies as found in transaction data. We compare this architecture to the previous approach and show in several experiments of varying complexity that our novel architecture provides better precision and stability.
We integrate this architecture into two generative neural network models adapted for transaction data and investigate how well normal behavior is modeled. We show that both architectures can successfully model normal transaction data, with our neural architecture improving generative performance for one model.
Since categorical and numerical variables are common in transaction data, but many machine learning methods only process numerical representations, we explore different representation learning techniques to transform categorical transaction data into dense numerical vectors. We extend this approach by proposing an outlier-aware discretization, thus incorporating numerical attributes into the computation of categorical embeddings, and investigate latent spaces, as well as quantitative performance for anomaly detection.
Next, we evaluate different scenarios for anomaly detection on transaction data. We extend our iNALU architecture to a neural layer that can model both numerical and non-numerical dependencies and evaluate it in a supervised and one-class setting. We investigate the stability and generalizability of our approach and show that it outperforms a variety of models in the balanced supervised setting and performs comparably in the one-class setting. Finally, we evaluate three approaches to using a generative model as an anomaly detector and compare the anomaly detection performance.
Neuropathies are a group of potentially treatable diseases with an often disabling and restricting course. Amyotrophic lateral sclerosis (ALS) is a lethal disease without causal treatment possibilities. The objective of this study was to examine the diagnostic utility of HRUS for the differentiation of subtypes of axonal and demyelinating neuropathies and to investigate its utility for the sonological differentiation of ALS.
The hypothetical statement that neuropathy causes enlargement of peripheral nerves compared to healthy controls proved to be right, but the adjunctive assumption that ALS does not cause enlargement of peripheral nerves proved to be wrong – in patients with ALS slight enlargement of peripheral nerves was visible as well. The statement that nerve enlargement can be detected by measurement of the cross-sectional area (CSA) and the longitudinal diameter (LD) with comparable results proved to be right, but the enlargement was slightly less present by measurement of the LD. The statement that axonal and demyelinating neuropathies show distinct patterns of nerve enlargement must be answered differentiated: The comparison between axonal and demyelinating neuropathies showed a stronger nerve enlargement in patients with demyelinating neuropathies than in patients with axonal neuropathies at proximal nerve segments of upper extremities. In the comparison of diagnose-defined subgroups of inflammatory demyelinating neuropathies a respective specific pattern of nerve enlargement was visible. However, remarkable in this context was the strong nerve enlargement found in patients with NSVN, which is classified as an axonal neuropathy. Stratification for specific findings in nerve biopsy did not lead to constructive differences in comparison between the different groups.
To sum up, HRUS showed to provide a useful contribution in the diagnostic process of neuropathies and ALS but needs to be integrated in a multimodal diagnostic approach.
In recent years, great progress has been made in the area of Artificial Intelligence (AI) due to the possibilities of Deep Learning which steadily yielded new state-of-the-art results especially in many image recognition tasks.
Currently, in some areas, human performance is achieved or already exceeded.
This great development already had an impact on the area of Optical Music Recognition (OMR) as several novel methods relying on Deep Learning succeeded in specific tasks.
Musicologists are interested in large-scale musical analysis and in publishing digital transcriptions in a collection enabling to develop tools for searching and data retrieving.
The application of OMR promises to simplify and thus speed-up the transcription process by either providing fully-automatic or semi-automatic approaches.
This thesis focuses on the automatic transcription of Medieval music with a focus on square notation which poses a challenging task due to complex layouts, highly varying handwritten notations, and degradation.
However, since handwritten music notations are quite complex to read, even for an experienced musicologist, it is to be expected that even with new techniques of OMR manual corrections are required to obtain the transcriptions.
This thesis presents several new approaches and open source software solutions for layout analysis and Automatic Text Recognition (ATR) for early documents and for OMR of Medieval manuscripts providing state-of-the-art technology.
Fully Convolutional Networks (FCN) are applied for the segmentation of historical manuscripts and early printed books, to detect staff lines, and to recognize neume notations.
The ATR engine Calamari is presented which allows for ATR of early prints and also the recognition of lyrics.
Configurable CNN/LSTM-network architectures which are trained with the segmentation-free CTC-loss are applied to the sequential recognition of text but also monophonic music.
Finally, a syllable-to-neume assignment algorithm is presented which represents the final step to obtain a complete transcription of the music.
The evaluations show that the performances of any algorithm is highly depending on the material at hand and the number of training instances.
The presented staff line detection correctly identifies staff lines and staves with an $F_1$-score of above $99.5\%$.
The symbol recognition yields a diplomatic Symbol Accuracy Rate (dSAR) of above $90\%$ by counting the number of correct predictions in the symbols sequence normalized by its length.
The ATR of lyrics achieved a Character Error Rate (CAR) (equivalently the number of correct predictions normalized by the sentence length) of above $93\%$ trained on 771 lyric lines of Medieval manuscripts and of 99.89\% when training on around 3.5 million lines of contemporary printed fonts.
The assignment of syllables and their corresponding neumes reached $F_1$-scores of up to $99.2\%$.
A direct comparison to previously published performances is difficult due to different materials and metrics.
However, estimations show that the reported values of this thesis exceed the state-of-the-art in the area of square notation.
A further goal of this thesis is to enable musicologists without technical background to apply the developed algorithms in a complete workflow by providing a user-friendly and comfortable Graphical User Interface (GUI) encapsulating the technical details.
For this purpose, this thesis presents the web-application OMMR4all.
Its fully-functional workflow includes the proposed state-of-the-art machine-learning algorithms and optionally allows for a manual intervention at any stage to correct the output preventing error propagation.
To simplify the manual (post-) correction, OMMR4all provides an overlay-editor that superimposes the annotations with a scan of the original manuscripts so that errors can easily be spotted.
The workflow is designed to be iteratively improvable by training better models as soon as new Ground Truth (GT) is available.
Recent advances in Natural Language Preprocessing (NLP) allow for a fully automatic extraction of character networks for an incoming text. These networks serve as a compact and easy to grasp representation of literary fiction. They offer an aggregated view of the text, which can be used during distant reading approaches for the analysis of literary hypotheses. In their core, the networks consist of nodes, which represent literary characters, and edges, which represent relations between characters. For an automatic extraction of such a network, the first step is the detection of the references of all fictional entities that are of importance for a text. References to the fictional entities appear in the form of names, noun phrases and pronouns and prior to this work, no components capable of automatic detection of character references were available. Existing tools are only capable of detecting proper nouns, a subset of all character references. When evaluated on the task of detecting proper nouns in the domain of literary fiction, they still underperform at an F1-score of just about 50%. This thesis uses techniques from the field of semi-supervised learning, such as Distant supervision and Generalized Expectations, and improves the results of an existing tool to about 82%, when evaluated on all three categories in literary fiction, but without the need for annotated data in the target domain. However, since this quality is still not sufficient, the decision to annotate DROC, a corpus comprising 90 fragments of German novels was made. This resulted in a new general purpose annotation environment titled as ATHEN, as well as annotated data that spans about 500.000 tokens in total. Using this data, the combination of supervised algorithms and a tailored rule based algorithm, which in combination are able to exploit both - local consistencies as well as global consistencies - yield an algorithm with an F1-score of about 93%. This component is referred to as the Kallimachos tagger.
A character network can not directly display references however, instead they need to be clustered so that all references that belong to a real world or fictional entity are grouped together. This process widely known as coreference resolution is a hard problem in the focus of research for more than half a century. This work experimented with adaptations of classical feature based machine learning, with a dedicated rule based algorithm and with modern techniques of Deep Learning, but no approach can surpass 55% B-Cubed F1, when evaluated on DROC. Due to this barrier, many researchers do not use a fully-fledged coreference resolution when they extract character networks, but only focus on a more forgiving subset- the names. For novels such as Alice's Adventures in Wonderland by Lewis Caroll, this would however only result in a network in which many important characters are missing. In order to integrate important characters into the network that are not named by the author, this work makes use of automatic detection of speaker and addressees for direct speech utterances (all entities involved in a dialog are considered to be of importance). This problem is by itself not an easy task, however the most successful system analysed in this thesis is able to correctly determine the speaker to about 85% of the utterances as well as about 65% of the addressees. This speaker information can not only help to identify the most dominant characters, but also serves as a way to model the relations between entities.
During the span of this work, components have been developed to model relations between characters using speaker attribution, using co-occurrences as well as by the usage of true interactions, for which yet again a dataset was annotated using ATHEN. Furthermore, since relations between characters are usually typed, a component for the extraction of a typed relation was developed. Similar to the experiments for the character reference detection, a combination of a rule based and a Maximum Entropy classifier yielded the best overall results, with the extraction of family relations showing a score of about 80% and the quality of love relations with a score of about 50%. For family relations, a kernel for a Support Vector Machine was developed that even exceeded the scores of the combined approach but is behind on the other labels.
In addition, this work presents new ways to evaluate automatically extracted networks without the need of domain experts, instead it relies on the usage of expert summaries. It also refrains from the uses of social network analysis for the evaluation, but instead presents ranked evaluations using Precision@k and the Spearman Rank correlation coefficient for the evaluation of the nodes and edges of the network. An analysis using these metrics showed, that the central characters of a novel are contained with high probability but the quality drops rather fast if more than five entities are analyzed. The quality of the edges is mainly dominated by the quality of the coreference resolution and the correlation coefficient between gold edges and system edges therefore varies between 30 and 60%.
All developed components are aggregated alongside a large set of other preprocessing modules in the Kallimachos pipeline and can be reused without any restrictions.
An Intelligent Semi-Automatic Workflow for Optical Character Recognition of Historical Printings
(2020)
Optical Character Recognition (OCR) on historical printings is a challenging task mainly due to the complexity of the layout and the highly variant typography. Nevertheless, in the last few years great progress has been made in the area of historical OCR resulting in several powerful open-source tools for preprocessing, layout analysis and segmentation, Automatic Text Recognition (ATR) and postcorrection. Their major drawback is that they only offer limited applicability by non-technical users like humanist scholars, in particular when it comes to the combined use of several tools in a workflow. Furthermore, depending on the material, these tools are usually not able to fully automatically achieve sufficiently low error rates, let alone perfect results, creating a demand for an interactive postcorrection functionality which, however, is generally not incorporated.
This thesis addresses these issues by presenting an open-source OCR software called OCR4all which combines state-of-the-art OCR components and continuous model training into a comprehensive workflow. While a variety of materials can already be processed fully automatically, books with more complex layouts require manual intervention by the users. This is mostly due to the fact that the required Ground Truth (GT) for training stronger mixed models (for segmentation as well as text recognition) is not available, yet, neither in the desired quantity nor quality.
To deal with this issue in the short run, OCR4all offers better recognition capabilities in combination with a very comfortable Graphical User Interface (GUI) that allows error corrections not only in the final output, but already in early stages to minimize error propagation. In the long run this constant manual correction produces large quantities of valuable, high quality training material which can be used to improve fully automatic approaches. Further on, extensive configuration capabilities are provided to set the degree of automation of the workflow and to make adaptations to the carefully selected default parameters for specific printings, if necessary. The architecture of OCR4all allows for an easy integration (or substitution) of newly developed tools for its main components by supporting standardized interfaces like PageXML, thus aiming at continual higher automation for historical printings.
In addition to OCR4all, several methodical extensions in the form of accuracy improving techniques for training and recognition are presented. Most notably an effective, sophisticated, and adaptable voting methodology using a single ATR engine, a pretraining procedure, and an Active Learning (AL) component are proposed. Experiments showed that combining pretraining and voting significantly improves the effectiveness of book-specific training, reducing the obtained Character Error Rates (CERs) by more than 50%.
The proposed extensions were further evaluated during two real world case studies: First, the voting and pretraining techniques are transferred to the task of constructing so-called mixed models which are trained on a variety of different fonts. This was done by using 19th century Fraktur script as an example, resulting in a considerable improvement over a variety of existing open-source and commercial engines and models. Second, the extension from ATR on raw text to the adjacent topic of typography recognition was successfully addressed by thoroughly indexing a historical lexicon that heavily relies on different font types in order to encode its complex semantic structure.
During the main experiments on very complex early printed books even users with minimal or no experience were able to not only comfortably deal with the challenges presented by the complex layout, but also to recognize the text with manageable effort and great quality, achieving excellent CERs below 0.5%. Furthermore, the fully automated application on 19th century novels showed that OCR4all (average CER of 0.85%) can considerably outperform the commercial state-of-the-art tool ABBYY Finereader (5.3%) on moderate layouts if suitably pretrained mixed ATR models are available.