The automatic cell segmentation and classification from whole slide images plays an important role in digital pathology, unlocking new opportunities for biomarker discovery. Despite extensive research, this task faces persistent challenges such as the differentiation of epithelial cells into normal and malignant. Many existing models lack reporting of epithelial subtyping, and when available, their performance is often suboptimal. This work benchmarks state-of-the-art methods to highlight this limitation and introduces GrEp, a geometric deep learning strategy that considers the broader epithelium tissue architecture to infer cell-level classification rather than relying exclusively on nuclei morphology. The proposed graph-based workflow significantly outperformed state-of-the-art nuclei classification models in colorectal cancer and generalized effectively to two unseen tissue types, endometrium and pancreas, proving the robustness of the geometry-based model. Given its speed and accuracy, we believe GrEp to be a valuable method to refine epithelial cell classification for downstream analyses in clinical and research settings.

Supervised learning is constrained by the availability of labeled data, which are especially expensive to acquire in the field of digital pathology. Making use of open-source data for pre-training or using domain adaptation can be a way to overcome this issue. However, pre-trained networks often fail to generalize to new test domains that are not distributed identically due to tissue stainings, types, and textures variations. Additionally, current domain adaptation methods mainly rely on fully-labeled source datasets. In this work, we propose Self-Rule to Multi-Adapt (SRMA), which takes advantage of self-supervised learning to perform domain adaptation, and removes the necessity of fully-labeled source datasets. SRMA can effectively transfer the discriminative knowledge obtained from a few labeled source domain's data to a new target domain without requiring additional tissue annotations. Our method harnesses both domains' structures by capturing visual similarity with intra-domain and cross-domain self-supervision. Moreover, we present a generalized formulation of our approach that allows the framework to learn from multiple source domains. We show that our proposed method outperforms baselines for domain adaptation of colorectal tissue type classification in single and multi-source settings, and further validate our approach on an in-house clinical cohort. The code and trained models are available open-source: https://github.com/christianabbet/SRA.

We propose to classify intestinal glands as normal or dysplastic using cell-graphs and graph-based deep learning methods. Dysplastic intestinal glands can lead to colorectal cancer, which is one of the three most common cancer types in the world. In order to assess the cancer stage and thus the treatment of a patient, pathologists analyse tissue samples of affected patients. Among other factors, they look at the changes in morphology of different tissues, such as the intestinal glands. Cell-graphs have a high representational power and can describe topological and geometrical properties of intestinal glands. However, classical graph-based methods have a high computational complexity and there is only a limited range of machine learning methods available. In this paper, we propose Graph Neural Networks (GNNs) as an efficient learning-based approach to classify cell-graphs. We investigate different variants of so-called Message Passing Neural Networks and compare them with a classical graph-based approach based on approximated Graph Edit Distance and k-nearest neighbours classifier. A promising classification accuracy of 94.8% is achieved by the proposed method on the pT1 Gland Graph dataset, which is an increase of 11.5% over the baseline result.

This paper presents SwissCrawl, the largest Swiss German text corpus to date. Composed of more than half a million sentences, it was generated using a customized web scraping tool that could be applied to other low-resource languages as well. The approach demonstrates how freely available web pages can be used to construct comprehensive text corpora, which are of fundamental importance for natural language processing. In an experimental evaluation, we show that using the new corpus leads to significant improvements for the task of language modeling.

In sequence modeling tasks the token order matters, but this information can be partially lost due to the discretization of the sequence into data points. In this paper, we study the imbalance between the way certain token pairs are included in data points and others are not. We denote this a token order imbalance (TOI) and we link the partial sequence information loss to a diminished performance of the system as a whole, both in text and speech processing tasks. We then provide a mechanism to leverage the full token order information—Alleviated TOI—by iteratively overlapping the token composition of data points. For recurrent networks, we use prime numbers for the batch size to avoid redundancies when building batches from overlapped data points. The proposed method achieved state of the art performance in both text and speech related tasks.

In recent years, different approaches for handwriting recognition that are based on graph representations have been proposed (e.g. graph-based keyword spotting or signature verification). This trend is mostly due to the availability of novel fast graph matching algorithms, as well as the inherent flexibility and expressivity of graph data structures when compared to vectorial representations. That is, graphs are able to directly adapt their size and structure to the size and complexity of the respective handwritten entities. However, the vast majority of the proposed approaches match the graphs from a global perspective only. In the present paper, we propose to match the underlying graphs from different local perspectives and combine the resulting assignments by means of Dynamic Time Warping. Moreover, we show that the proposed approach can be readily combined with global matchings. In an experimental evaluation, we employ the novel method in a signature verification scenario on two widely used benchmark datasets. On both datasets, we empirically confirm that the proposed approach outperforms state-of-the-art methods with respect to both accuracy and runtime.

Automatic analysis of scanned historical documents comprises a wide range of image analysis tasks, which are often challenging for machine learning due to a lack of human-annotated learning samples. With the advent of deep neural networks, a promising way to cope with the lack of training data is to pre-train models on images from a different domain and then fine-tune them on historical documents. In the current research, a typical example of such cross-domain transfer learning is the use of neural networks that have been pre-trained on the ImageNet database for object recognition. It remains a mostly open question whether or not this pre-training helps to analyse historical documents, which have fundamentally different image properties when compared with ImageNet. In this paper, we present a comprehensive empirical survey on the effect of ImageNet pre-training for diverse historical document analysis tasks, including character recognition, style classification, manuscript dating, semantic segmentation, and content-based retrieval. While we obtain mixed results for semantic segmentation at pixel-level, we observe a clear trend across different network architectures that ImageNet pre-training has a positive effect on classification as well as content-based retrieval.

Scanned handwritten historical documents are often not well accessible due to the limited feasibility of automatic full transcriptions. Thus, Keyword Spotting (KWS) has been proposed as an alternative to retrieve arbitrary query words from this kind of documents. In the present paper, word images are represented by means of graphs. That is, a graph is used to represent the inherent topological characteristics of handwriting. The actual keyword spotting is then based on matching a query graph with all document graphs. In particular, we make use of a fast graph matching algorithm that considers the contextual substructure of nodes. The motivation for this inclusion of node context is to increase the overall KWS accuracy. In an experimental evaluation on four historical documents, we show that the proposed procedure clearly outperforms diverse other template-based reference systems. Moreover, our novel framework keeps up or even outperforms many state-of-the-art learning-based KWS approaches.

For handwritten signature verification, signature images are typically represented with fixed-sized feature vectors capturing local and global properties of the handwriting. Graph-based representations offer a promising alternative, as they are flexible in size and model the global structure of the handwriting. However, they are only rarely used for signature verification, which may be due to the high computational complexity involved when matching two graphs. In this paper, we take a closer look at two recently presented structural methods for handwriting analysis, for which efficient matching methods are available: keypoint graphs with approximate graph edit distance and inkball models. Inkball models, in particular, have never been used for signature verification before. We investigate both approaches individually and propose a combined verification system, which demonstrates an excellent performance on the MCYT and GPDS benchmark data sets when compared with the state of the art.

Graph representations have been widely used in pattern recognition thanks to their powerful representation formalism and rich theoretical background. A number of error-tolerant graph matching algorithms such as graph edit distance have been proposed for computing a distance between two labelled graphs. However, they typically suffer from a high computational complexity, which makes it difficult to apply these matching algorithms in a real scenario. In this paper, we propose an efficient graph distance based on the emerging field of geometric deep learning. Our method employs a message passing neural network to capture the graph structure and learns a metric with a siamese network approach. The performance of the proposed graph distance is validated in two application cases, graph classification and graph retrieval of handwritten words, and shows a promising performance when compared with (approximate) graph edit distance benchmarks.

The dynamic signature is a biometric trait widely used and accepted for verifying a person's identity. Current automatic signature-based biometric systems typically require five, ten, or even more specimens of a person's signature to learn intrapersonal variability sufficient to provide an accurate verification of the individual's identity. To mitigate this drawback, this paper proposes a procedure for training with only a single reference signature. Our strategy consists of duplicating the given signature a number of times and training an automatic signature verifier with each of the resulting signatures. The duplication scheme is based on a sigma lognormal decomposition of the reference signature. Two methods are presented to create human-like duplicated signatures: the first varies the strokes' lognormal parameters (stroke-wise) whereas the second modifies their virtual target points (target-wise). A challenging benchmark, assessed with multiple state-of-the-art automatic signature verifiers and multiple databases, proves the robustness of the system. Experimental results suggest that our system, with a single reference signature, is capable of achieving a similar performance to standard verifiers trained with up to five signature specimens.

Graphs provide a powerful representation formalism for handwritten signatures, capturing local properties as well as their relations. Yet, although introduced early for signature verification, only a few current systems rely on graph-based representations. A possible reason is the high computational complexity involved for matching two general graphs. In this paper, we introduce a novel structural approach to offline signature verification using an efficient cubic-time approximation of graph edit distance. We put forward several ways of creating, normalizing, and comparing signature graphs built from keypoints and investigate their performance on three benchmark datasets. The experiments demonstrate a promising performance of the proposed structural approach when compared with the state of the art.

Keyword Spotting (KWS) offers a convenient way to improve the accessibility to historical handwritten documents by retrieving search terms in scanned document images. The approach for KWS proposed in the present paper is based on segmented word images that are represented by means of different types of graphs. The actual keyword spotting is based on matching a query graph with a set of document graphs using the concept of graph edit distance. In particular, we propose to employ ensemble methods for KWS with graphs. That is, a query graph is not matched against one but several different graphs representing the same document word. Eventually, we use different strategies to combine these individual graph dissimilarities. In an experimental evaluation on two benchmark datasets, the proposed ensemble methods outperform the individual ensemble members as well as four state-of-the-art reference systems based on dynamic time warping.

When using tablet computers, smartphones, or digital pens, human users perform movements with a stylus or their fingers that can be analyzed by the kinematic theory of rapid human movements. In this paper, we present a user-centered system for signature verification that performs such a kinematic analysis to verify the identity of the user. It is one of the first systems that is based on a direct comparison of the elementary neuromuscular strokes which are detected in the handwriting. Taking into account the number of strokes, their similarity, and their timing, the string edit distance is employed to derive a dissimilarity measure for signature verification. On several benchmark datasets, we demonstrate that this neuromuscular analysis is complementary to a well-established verification using dynamic time warping. By combining both approaches, our verifier is able to outperform current state-of-the-art results in on-line signature verification.

In historical manuscripts, humans can detect handwritten words, lines, and decorations with lightness even if they do not know the language or the script. Yet for automatic processing this task has proven elusive, especially in the case of handwritten documents with complex layouts, which is why semiautomatic methods that integrate the human user into the process are needed. In this paper, we introduce a user-centered segmentation method based on document graphs and scribbling interaction. The graphs capture a sparse representation of the document's structure that can then be edited by the user with a stylus on a touch-sensitive screen. We evaluate the proposed method on a newly introduced database of historical manuscripts with complex layout and demonstrate, first, that the document graphs are already close to the desired segmentation and, second, that scribbling allows a natural and efficient interaction.

Ground truth is both - indispensable for training and evaluating document analysis methods, and yet very tedious to create manually. This especially holds true for complex historical manuscripts that exhibit challenging layouts with interfering and overlapping handwriting. In this paper, we propose a novel semi-automatic system to support layout annotations in such a scenario based on document graphs and a pen-based scribbling interaction. On the one hand, document graphs provide a sparse page representation that is already close to the desired ground truth and on the other hand, scribbling facilitates an efficient and convenient pen-based interaction with the graph. The performance of the system is demonstrated in the context of a newly introduced database of historical manuscripts with complex layouts.

Inkball models provide a tool for matching and comparison of spatially structured markings such as handwritten characters and words. Hidden Markov models offer a framework for decoding a stream of text in terms of the most likely sequence of causal states. Prior work with HMM has relied on observation of features that are correlated with underlying characters, without modeling them directly. This paper proposes to use the results of inkball-based character matching as a feature set input directly to the HMM. Experiments indicate that this technique outperforms other tested methods at handwritten word recognition on a common benchmark when applied without normalization or text deslanting.

Deep learning had a significant impact on diverse pattern recognition tasks in the recent past. In this paper, we investigate its potential for keyword spotting in handwritten documents by designing a novel feature extraction system based on Convolutional Deep Belief Networks. Sliding window features are learned from word images in an unsupervised manner. The proposed features are evaluated both for template-based word spotting with Dynamic Time Warping and for learning-based word spotting with Hidden Markov Models. In an experimental evaluation on three benchmark data sets with historical and modern handwriting, it is shown that the proposed learned features outperform three standard sets of handcrafted features.

What can be done with only one enrolled real hand-written signature in Automatic Signature Verification (ASV)? Using 5 or 10 signatures for training is the most common case to evaluate ASV. In the scarcely addressed case of only one available signature for training, we propose to use modified duplicates. Our novel technique relies on a fully neuromuscular representation of the signatures based on the Kinematic Theory of rapid human movements and its Sigma-Lognormal model. This way, a real on-line signature is converted into the Sigma-Lognormal model domain. The model parameters are then varied to generate new duplicated signatures.

In the field of automatic signature verification, a major challenge for statistical analysis and pattern recognition is the small number of reference signatures per user. Score normalization, in particular, is challenged by the lack of information about intra-user variability. In this paper, we analyze several approaches to score normalization for dynamic time warping and propose a new two-stage normalization which detects simple forgeries in a first stage and copes with more skilled forgeries in a second stage. An experimental evaluation is conducted on two data sets with different characteristics, namely the MCYT online signature corpus, which contains over three hundred users, and the SUSIG visual sub-corpus, which contains highly skilled forgeries. The results demonstrate that score normalization is a key component for signature verification and that the proposed two-stage normalization achieves some of the best results on these difficult data sets both for random and for skilled forgeries.

The concept of graph edit distance (GED) is still one of the most flexible and powerful graph matching approaches available. Yet, exact computation of GED can be solved in exponential time complexity only. A previously introduced approximation framework reduces the computation of GED to an instance of a linear sum assignment problem. Major benefit of this reduction is that an optimal assignment of nodes (including local structures) can be computed in polynomial time. Given this assignment an approximate value of GED can be immediately derived. Yet, this approach considers local — rather than the global — structural properties of the graphs only, and thus GED derived from the optimal node assignment generally overestimates the true edit distance. Recently, it has been shown how the existing approximation framework can be exploited to additionally derive a lower bound of the exact edit distance without any additional computations. In this paper we make use of regression analysis in order to predict the exact GED using these two bounds. In an experimental evaluation on diverse graph data sets we empirically verify the gain of distance accuracy of the estimated GEDs compared to both bounds.

Automated reading of historical handwriting is needed to search and browse ancient manuscripts in digital libraries based on their textual content. In this paper, we present a combined system for text localization and transcription in page images. It includes flexible learning-based methods for layout analysis and handwriting recognition, which were developed in the context of the Swiss research project HisDoc. A comprehensive experimental evaluation is provided for the medieval Parzival database, demonstrating a promising word recognition accuracy of 93.0% with closed vocabulary. In order to harmonize the evaluation of the two document analysis tasks, we introduce a novel evaluation measure for text line extraction that takes substitution, deletion, as well as insertion errors into account.

With increasing computational power, the trend in unconstrained text recognition is going towards whole document processing. For this task, more sophisticated language models can be employed. One approach is to take advantage the fact that the text of a document normally deals with a specific topic and hence the word occurrence probability is biased. Cache language models combine information about recent words, the cache, with a general statistical language model to increase the recognition rate. In this work we introduce a modified version of the cache language model to the task of handwriting recognition, where the N-best recognition output of the entire document is used to refine the language model for a consecutive recognition pass. An experimental evaluation on the IAM database demonstrates that the word error rate can be reduced with the proposed cache language model.

The use of Higher-Order Singular Value Decomposition (HOSVD) and other tensor decomposition methods are popular in the face recognition domain, yet a direct application to handwritten character recognition has not shown promising results so far. Character recognition is commonly performed in two steps: feature extraction and classification. In this paper, we propose a feature extraction algorithm based on HOSVD which is then combined with standard statistical classification. The algorithm constructs a tensor from the training data and applies HOSVD in order to obtain a feature extractor matrix for arbitrary character images. We evaluate the proposed handwriting features in combination with SVM classification for character recognition on the CEDAR benchmark data set. The results indicate that our proposed approach significantly outperforms the standard HOSVD classification method.

A fully automatic framework has been introduced recently for neuromuscular representation of complex handwriting patterns, such as gestures, signatures, and words, based on the Kinematic Theory of rapid human movements and its Sigma-Lognormal model. In this paper, we investigate the application of this framework to unconstrained whiteboard notes, taking into account a novel acquisition modality, multiple writers, natural language, and complete text lines. Although these conditions deviate strongly from the previously considered scenario of brief pen movements on tablet computers, we demonstrate that the Sigma-Lognormal model is still able to represent the handwriting accurately. In order to deal with longer handwriting patterns, we propose a robust component-wise representation of text lines that achieves a high model quality. Furthermore, we propose a stroke-wise distortion method to generate synthetic text lines from the Sigma-Lognormal representation of real specimens. For handwriting recognition on the IAM online database, it is demonstrated that the extension of the training set with the proposed synthesis method significantly increases current benchmark results achieved with recurrent neural networks.

In recent years the authors of the present paper introduced a powerful approximation fra or the graph edit distance problem. The basic idea of this approximation is to build a square cost matrix C = (cj), where each entry reflects the cost of a node substitution, deletion or insertion plus the matching cost arising from the local edge structure. Based on C an optimal assignment of the nodes and their local structure can be established in polynomial time (using, for instance, the Hungarian algorithm). Since this approach considers the local -- rather than the global -- structural properties of the graphs only, the obtained graph edit distance value is suboptimal in the sense of overestimating the true edit distance in general. The present paper pursues the idea of including topological information in the node labels in order to increase the amount of structural information available during the initial assignment process. In an experimental evaluation on three real world data sets a reduction of the overestimation can be observed while the run time is only moderately increased compared to our original framework.

Facing high error rates and slow recognition speed for full text transcription of unconstrained handwriting images, keyword spotting is a promising alternative to locate specific search terms within scanned document images. We have previously proposed a learning-based method for keyword spotting using character hidden Markov models that showed a high performance when compared with traditional template image matching. In the lexicon-free approach pursued, only the text appearance was taken into account for recognition. In this paper, we integrate character n-gram language models into the spotting system in order to provide an additional language context. On the modern IAM database as well as the historical George Washington database, we demonstrate that character language models significantly improve the spotting performance.

Text line segmentation is one of the main parts of document image analysis, it provides crucial information for automated reading, word spotting, alignment between image and transcription, or indexing of documents. Yet it remains an open problem for handwritten historical documents because of complex layouts on the one hand, such as curved and touching text lines, and binarization problems on the other hand, caused by ornaments, wrinkles, stains, holes, etc. In this paper, we propose a binarization-free clustering method for text line segmentation that is not only able to cope with touching text lines, but also with complex baseline curvature. Avoiding the assumption of straight baselines, small interest point clusters are grouped into text lines based on their local orientation. Experiments conducted on artificially distorted images of the Saint Gall database show promising results.

Unconstrained on-line handwriting recognition is typically approached within the framework of generative HMM-based classifiers. In this paper, we introduce a novel discriminative method that relies, in contrast, on explicit grapheme segmentation and SVM-based character recognition. In addition to single character recognition with rejection, bi-characters are recognized in order to refine the recognition hypotheses. In particular, bi-character recognition is able to cope with the problem of shared character parts. Whole word recognition is achieved with an efficient dynamic programming method similar to the Viterbi algorithm. In an experimental evaluation on the Unipen-ICROW-03 database, we demonstrate improvements in recognition accuracy of up to 8% for a lexicon of 20,000 words with the proposed method when compared with an HMM-based baseline system. The computational speed is on par with the baseline system.

Segmenting page images into text lines is a crucial pre-processing step for automated reading of historical documents. Challenging issues in this open research field are given \eg by paper or parchment background noise, ink bleed-through, artifacts due to aging, stains, and touching text lines. In this paper, we present a novel binarization-free line segmentation method that is robust to noise and copes with overlapping and touching text lines. First, interest points representing parts of characters are extracted from gray-scale images. Next, word clusters are identified in high-density regions and touching components such as ascenders and descenders are separated using seam carving. Finally, text lines are generated by concatenating neighboring word clusters, where neighborhood is defined by the prevailing orientation of the words in the document. An experimental evaluation on the Latin manuscript images of the Saint Gall database shows promising results for real-world applications in terms of both accuracy and efficiency.

State-of-the-art handwriting recognition systems are learning-based systems that require large sets of training data. The creation of training data, and consequently the creation of a well-performing recognition system, requires therefore a substantial amount of human work. This can be reduced with semi-supervised learning, which uses unlabeled text lines for training as well. Current approaches estimate the correct transcription of the unlabeled data via handwriting recognition which is not only extremely demanding as far as computational costs are concerned but also requires a good model of the target language. In this paper, we propose a different approach that makes use of keyword spotting, which is significantly faster and does not need any language model. In a set of experiments we demonstrate its superiority over existing approaches.

Keyword spotting refers to the process of retrieving all instances of a given keyword from a document. In the present paper, a novel keyword spotting method for handwritten documents is described. It is derived from a neural network-based system for unconstrained handwriting recognition. As such it performs template-free spotting, i.e., it is not necessary for a keyword to appear in the training set. The keyword spotting is done using a modification of the CTC Token Passing algorithm in conjunction with a recurrent neural network. We demonstrate that the proposed systems outperform not only a classical dynamic time warping-based approach but also a modern keyword spotting system, based on hidden Markov models. Furthermore, we analyze the performance of the underlying neural networks when using them in a recognition task followed by keyword spotting on the produced transcription. We point out the advantages of keyword spotting when compared to classic text line recognition.

For historical documents, available transcriptions typically are inaccurate when compared with the scanned document images. Not only the position of the words and sentences are unknown, but also the correct image transcription may not be matched exactly. An error-tolerant alignment is needed to make the document images amenable to browsing and searching in digital libraries. In this paper, we propose a novel multi-pass alignment method based on Hidden Markov Models (HMM) that combines text line recognition, string alignment, and keyword spotting to cope with word substitutions, deletions, and insertions in the transcription. In a segmentation-free approach, transcriptions of complete pages are aligned with sequences of text line images. On the Parzival data set, results are reported for several degrees of artificial distortions. Both the accuracy and the efficiency of the proposed system are promising for real-world applications.

To cope with the tremendous variations of writing styles encountered between different individuals, unconstrained automatic handwriting recognition systems need to be trained on large sets of labeled data. Traditionally, the training data has to be labeled manually, which is a laborious and costly process. Semi-supervised learning techniques offer methods to utilize unlabeled data, which can be obtained cheaply in large amounts in order, to reduce the need for labeled data. In this paper, we propose the use of Co-Training for improving the recognition accuracy of two weakly trained handwriting recognition systems. The first one is based on Recurrent Neural Networks while the second one is based on Hidden Markov Models. On the IAM off-line handwriting database we demonstrate a significant increase of the recognition accuracy can be achieved with Co-Training for single word recognition.

Spotting keywords in handwritten documents without transcription is a valuable method as it allows one to search, index, and classify such documents. In this paper we show that keyword spotting based on bi-directional Long Short-Term Memory (BLSTM) recurrent neural nets can successfully be applied on online handwritten documents with non-text content. It even works without preprocessing steps such as text vs. non-text distinction and text line extraction. We also propose a modification that can improve the precision with little effort.

Handwriting recognition in historical documents is vital for the creation of digital libraries. The creation of readily available ground truth data plays a central role for the development of new recognition technologies. For historical documents, ground truth creation is more difficult and time-consuming when compared with modern documents. In this paper, we present a semi-automatic ground truth creation proceeding for historical documents that takes into account noisy background and transcription alignment. The proposed ground truth creation is demonstrated for the IAM Historical Handwriting Database (IAM-HistDB) that is currently under construction and will include several hundred Old German manuscripts. With a small set of algorithmic tools and few manual interactions, it is shown how laypersons can efficiently create a ground truth for handwriting recognition.

Automatic transcription of historical documents is vital for the creation of digital libraries. In this paper we propose graph similarity features as a novel descriptor for handwriting recognition in historical documents based on Hidden Markov Models. Using a structural graph-based representation of text images, a sequence of graph similarity features is extracted by means of dissimilarity embedding with respect to a set of character prototypes. On the medieval Parzival data set it is demonstrated that the proposed structural descriptor significantly outperforms two well-known statistical reference descriptors for single word recognition.

Being able to search for words or phrases in historic handwritten documents is of paramount importance when preserving cultural heritage. Storing scanned pages of written text can save the information from degradation, but it does not make the textual information readily available. Automatic keyword spotting systems for handwritten historic documents can fill this gap. However, most such systems have trouble with the great variety of writing styles. It is not uncommon for handwriting processing systems to be built for just a single book. In this paper we show that neural network based keyword spotting systems are flexible enough to be used successfully on historic data, even when they are trained on a modern handwriting database. We demonstrate that with little transcribed historic text, added to the training set, the performance can further be enhanced.

Handwritten word spotting aims at making document images amenable to browsing and searching by keyword retrieval. In this paper, we present a word spotting system based on Hidden Markov Models (HMM) that uses trained subword models to spot keywords. With the proposed method, arbitrary keywords can be spotted that do not need to be present in the training set. Also, no text line segmentation is required. On the modern IAM off-line database and the historical George Washington database we show that the proposed system outperforms a standard template matching approach based on dynamic time warping (DTW).

Building recognition systems for historical documents is a difficult task. Especially, when it comes to medieval scripts. The complexity is mainly affected by the poor quality and the small quantity of the data available. In this paper we apply an HMM based recognition system to medieval manuscripts from the 13th century written in Middle High German. The recognition system, which was originally developed for modern scripts, has been adapted to medieval scripts. Beside the data processing, one of the major challenges is to create a suitable language model. Because of the lack of appropriate independent text corpora for medieval languages, the language model has to be created on the base of a rather small number of manuscripts only. Due to the small size of the corpus, optimizing the language model parameters can quickly lead to the problem of overfitting. In this paper we describe a strategy to integrate all available information into the language model and to optimize the language model parameters without suffering from this problem.

In structural pattern recognition, a major drawback of graph based representation is the lack of algorithmic tools. To overcome this lack, we embed graphs in vector spaces by means of prototype selection and graph edit distance, thus making them available to all algorithms of statistical pattern recognition that operate on feature vectors. In previous work a similar procedure was applied. However, the only classifier used within this framework was support vector machine (SVM). In the present paper, we significantly extend the scope of the previous work and present an experimental study where, in addition to SVM, a number of other well established classifiers from statistical pattern recognition are used for graph classification. On a total of five different graph data sets of diverse nature it is demonstrated that the proposed graph embedding in conjunction with standard classifiers from statistical pattern recognition has great potential to outperform classification methods applied in the original graph domain.