Mathias Seuret

Konstantina Nikolaidou, George Retsinas, Vincent Christlein et al.

Text-to-Image synthesis is the task of generating an image according to a specific text description. Generative Adversarial Networks have been considered the standard method for image synthesis virtually since their introduction. Denoising Diffusion Probabilistic Models are recently setting a new baseline, with remarkable results in Text-to-Image synthesis, among other fields. Aside its usefulness per se, it can also be particularly relevant as a tool for data augmentation to aid training models for other document image processing tasks. In this work, we present a latent diffusion-based method for styled text-to-text-content-image generation on word-level. Our proposed method is able to generate realistic word image samples from different writer styles, by using class index styles and text content prompts without the need of adversarial training, writer recognition, or text recognition. We gauge system performance with the Fréchet Inception Distance, writer recognition accuracy, and writer retrieval. We show that the proposed model produces samples that are aesthetically pleasing, help boosting text recognition performance, and get similar writer retrieval score as real data. Code is available at: https://github.com/koninik/WordStylist.

Marco Peer, Thomas Gorges, Mathias Seuret et al.

Popular initiatives and referendums are central to Swiss democracy, yet the validation of handwritten signature lists remains a labor-intensive manual process. This paper investigates the potential of automated document analysis methods, including OCR and AI-based handwriting analysis, to support this task. We propose a pipeline combining template-based line segmentation with text recognition and writer retrieval techniques, evaluated on a dataset of 443 handwritten entries from 418 writers. Results show that OCR struggles with out-of-vocabulary handwriting, with a CER of 29.6% for first names. In contrast, writer retrieval performs more robustly, reaching an mAP of 50.6%. Furthermore, our experiments indicate that off-the-shelf OCR systems are not sufficiently reliable for transcription of handwritten signature data, particularly for short, out-of-vocabulary entries such as names or addresses. However, writer retrieval methods can effectively identify visually similar entries across signature lists, making them a suitable tool for supporting the detection of potential duplicate submissions based on handwriting similarity.

Konstantina Nikolaidou, Mathias Seuret, Hamam Mokayed et al.

This paper presents a systematic literature review of image datasets for document image analysis, focusing on historical documents, such as handwritten manuscripts and early prints. Finding appropriate datasets for historical document analysis is a crucial prerequisite to facilitate research using different machine learning algorithms. However, because of the very large variety of the actual data (e.g., scripts, tasks, dates, support systems, and amount of deterioration), the different formats for data and label representation, and the different evaluation processes and benchmarks, finding appropriate datasets is a difficult task. This work fills this gap, presenting a meta-study on existing datasets. After a systematic selection process (according to PRISMA guidelines), we select 65 studies that are chosen based on different factors, such as the year of publication, number of methods implemented in the article, reliability of the chosen algorithms, dataset size, and journal outlet. We summarize each study by assigning it to one of three pre-defined tasks: document classification, layout structure, or content analysis. We present the statistics, document type, language, tasks, input visual aspects, and ground truth information for every dataset. In addition, we provide the benchmark tasks and results from these papers or recent competitions. We further discuss gaps and challenges in this domain. We advocate for providing conversion tools to common formats (e.g., COCO format for computer vision tasks) and always providing a set of evaluation metrics, instead of just one, to make results comparable across studies.

Anguelos Nicolaou, Vincent Christlein, Edgar Riba et al.

We propose the use of fractals as a means of efficient data augmentation. Specifically, we employ plasma fractals for adapting global image augmentation transformations into continuous local transforms. We formulate the diamond square algorithm as a cascade of simple convolution operations allowing efficient computation of plasma fractals on the GPU. We present the TorMentor image augmentation framework that is totally modular and deterministic across images and point-clouds. All image augmentation operations can be combined through pipelining and random branching to form flow networks of arbitrary width and depth. We demonstrate the efficiency of the proposed approach with experiments on document image segmentation (binarization) with the DIBCO datasets. The proposed approach demonstrates superior performance to traditional image augmentation techniques. Finally, we use extended synthetic binary text images in a self-supervision regiment and outperform the same model when trained with limited data and simple extensions.

Vincent Christlein, Isabelle Marthot-Santaniello, Martin Mayr et al.

The analysis of digitized historical manuscripts is typically addressed by paleographic experts. Writer identification refers to the classification of known writers while writer retrieval seeks to find the writer by means of image similarity in a dataset of images. While automatic writer identification/retrieval methods already provide promising results for many historical document types, papyri data is very challenging due to the fiber structures and severe artifacts. Thus, an important step for an improved writer identification is the preprocessing and feature sampling process. We investigate several methods and show that a good binarization is key to an improved writer identification in papyri writings. We focus mainly on writer retrieval using unsupervised feature methods based on traditional or self-supervised-based methods. It is, however, also comparable to the state of the art supervised deep learning-based method in the case of writer classification/re-identification.

Martin Mayr, Marcel Dreier, Florian Kordon et al.

The imitation of cursive handwriting is mainly limited to generating handwritten words or lines. Multiple synthetic outputs must be stitched together to create paragraphs or whole pages, whereby consistency and layout information are lost. To close this gap, we propose a method for imitating handwriting at the paragraph level that also works for unseen writing styles. Therefore, we introduce a modified latent diffusion model that enriches the encoder-decoder mechanism with specialized loss functions that explicitly preserve the style and content. We enhance the attention mechanism of the diffusion model with adaptive 2D positional encoding and the conditioning mechanism to work with two modalities simultaneously: a style image and the target text. This significantly improves the realism of the generated handwriting. Our approach sets a new benchmark in our comprehensive evaluation. It outperforms all existing imitation methods at both line and paragraph levels, considering combined style and content preservation.

Changhun Kim, Martin Mayr, Thomas Gorges et al.

State-of-the-art handwritten text recognition (HTR) systems commonly use Transformers, whose growing key-value (KV) cache makes decoding slow and memory-intensive. We introduce DRetHTR, a decoder-only model built on Retentive Networks (RetNet). Compared to an equally sized decoder-only Transformer baseline, DRetHTR delivers 1.6-1.9x faster inference with 38-42% less memory usage, without loss of accuracy. By replacing softmax attention with softmax-free retention and injecting multi-scale sequential priors, DRetHTR avoids a growing KV cache: decoding is linear in output length in both time and memory. To recover the local-to-global inductive bias of attention, we propose layer-wise gamma scaling, which progressively enlarges the effective retention horizon in deeper layers. This encourages early layers to model short-range dependencies and later layers to capture broader context, mitigating the flexibility gap introduced by removing softmax. Consequently, DRetHTR achieves best reported test character error rates of 2.26% (IAM-A, en), 1.81% (RIMES, fr), and 3.46% (Bentham, en), and is competitive on READ-2016 (de) with 4.21%. This demonstrates that decoder-only RetNet enables Transformer-level HTR accuracy with substantially improved decoding speed and memory efficiency.

Richin Sukesh, Mathias Seuret, Anguelos Nicolaou et al.

Binarization of document images is an important pre-processing step in the field of document analysis. Traditional image binarization techniques usually rely on histograms or local statistics to identify a valid threshold to differentiate between different aspects of the image. Deep learning techniques are able to generate binarized versions of the images by learning context-dependent features that are less error-prone to degradation typically occurring in document images. In recent years, many deep learning-based methods have been developed for document binarization. But which one to choose? There have been no studies that compare these methods rigorously. Therefore, this work focuses on the evaluation of different deep learning-based methods under the same evaluation protocol. We evaluate them on different Document Image Binarization Contest (DIBCO) datasets and obtain very heterogeneous results. We show that the DE-GAN model was able to perform better compared to other models when evaluated on the DIBCO2013 dataset while DP-LinkNet performed best on the DIBCO2017 dataset. The 2-StageGAN performed best on the DIBCO2018 dataset while SauvolaNet outperformed the others on the DIBCO2019 challenge. Finally, we make the code, all models and evaluation publicly available (https://github.com/RichSu95/Document_Binarization_Collection) to ensure reproducibility and simplify future binarization evaluations.

Michele Alberti, Angela Botros, Narayan Schuez et al.

In this work, we investigate the application of trainable and spectrally initializable matrix transformations on the feature maps produced by convolution operations. While previous literature has already demonstrated the possibility of adding static spectral transformations as feature processors, our focus is on more general trainable transforms. We study the transforms in various architectural configurations on four datasets of different nature: from medical (ColorectalHist, HAM10000) and natural (Flowers, ImageNet) images to historical documents (CB55) and handwriting recognition (GPDS). With rigorous experiments that control for the number of parameters and randomness, we show that networks utilizing the introduced matrix transformations outperform vanilla neural networks. The observed accuracy increases by an average of 2.2 across all datasets. In addition, we show that the benefit of spectral initialization leads to significantly faster convergence, as opposed to randomly initialized matrix transformations. The transformations are implemented as auto-differentiable PyTorch modules that can be incorporated into any neural network architecture. The entire code base is open-source.

Thomas Gorges, Janne van der Loop, Lukas Hüttner et al.

This paper presents CircleID, a large-scale ICDAR 2026 competition on writer identification and pen classification from scanned hand-drawn circles. The primary objective is to investigate how biometric writer characteristics and physical pen features naturally entangle within minimal, static traces. CircleID comprises two distinct tasks: (1) open-set writer identification, requiring models to recognize known writers while explicitly rejecting unknown ones, and (2) cross-writer pen classification, evaluated across both seen and unseen writers. Participants were provided with a new, controlled dataset of 46,155 tightly cropped circle images, digitized at 400 DPI and annotated for writer identity and pen type. The dataset comprises samples from 50 known and 16 unknown writers using eight different pens. Hosted on Kaggle as two separate tracks with public and private leaderboards, the competition provided participants with a ResNet baseline. In total, 389 teams (436 participants) made 3,185 submissions for the pen classification task, and 113 teams (141 participants) made 1,737 submissions for the writer identification track. The best-performing private leaderboard submissions achieved a Top-1 accuracy of 64.801% for writer identification and 92.726% for pen classification. This paper details the dataset, evaluates the winning methodologies, and analyzes the impact of out-of-distribution writers on model generalization and feature disentanglement. In this large-scale competition, CircleID establishes a new baseline for minimal-trace analysis.

Mathias Seuret, Janne van der Loop, Nikolaus Weichselbaumer et al.

In this paper, we investigate the usage of fine-grained font recognition on OCR for books printed from the 15th to the 18th century. We used a newly created dataset for OCR of early printed books for which fonts are labeled with bounding boxes. We know not only the font group used for each character, but the locations of font changes as well. In books of this period, we frequently find font group changes mid-line or even mid-word that indicate changes in language. We consider 8 different font groups present in our corpus and investigate 13 different subsets: the whole dataset and text lines with a single font, multiple fonts, Roman fonts, Gothic fonts, and each of the considered fonts, respectively. We show that OCR performance is strongly impacted by font style and that selecting fine-tuned models with font group recognition has a very positive impact on the results. Moreover, we developed a system using local font group recognition in order to combine the output of multiple font recognition models, and show that while slower, this approach performs better not only on text lines composed of multiple fonts but on the ones containing a single font only as well.

Alexander Mattick, Martin Mayr, Mathias Seuret et al.

As of recent generative adversarial networks have allowed for big leaps in the realism of generated images in diverse domains, not the least of which being handwritten text generation. The generation of realistic-looking hand-written text is important because it can be used for data augmentation in handwritten text recognition (HTR) systems or human-computer interaction. We propose SmartPatch, a new technique increasing the performance of current state-of-the-art methods by augmenting the training feedback with a tailored solution to mitigate pen-level artifacts. We combine the well-known patch loss with information gathered from the parallel trained handwritten text recognition system and the separate characters of the word. This leads to a more enhanced local discriminator and results in more realistic and higher-quality generated handwritten words.

Mathias Seuret, Anguelos Nicolaou, Dominique Stutzmann et al.

This competition succeeds upon a line of competitions for writer and style analysis of historical document images. In particular, we investigate the performance of large-scale retrieval of historical document fragments in terms of style and writer identification. The analysis of historic fragments is a difficult challenge commonly solved by trained humanists. In comparison to previous competitions, we make the results more meaningful by addressing the issue of sample granularity and moving from writer to page fragment retrieval. The two approaches, style and author identification, provide information on what kind of information each method makes better use of and indirectly contribute to the interpretability of the participating method. Therefore, we created a large dataset consisting of more than 120 000 fragments. Although the most teams submitted methods based on convolutional neural networks, the winning entry achieves an mAP below 40%.

Martin Leipert, Georg Vogeler, Mathias Seuret et al.

Notarial instruments are a category of documents. A notarial instrument can be distinguished from other documents by its notary sign, a prominent symbol in the certificate, which also allows to identify the document's issuer. Naturally, notarial instruments are underrepresented in regard to other documents. This makes a classification difficult because class imbalance in training data worsens the performance of Convolutional Neural Networks. In this work, we evaluate different countermeasures for this problem. They are applied to a binary classification and a segmentation task on a collection of medieval documents. In classification, notarial instruments are distinguished from other documents, while the notary sign is separated from the certificate in the segmentation task. We evaluate different techniques, such as data augmentation, under- and oversampling, as well as regularizing with focal loss. The combination of random minority oversampling and data augmentation leads to the best performance. In segmentation, we evaluate three loss-functions and their combinations, where only class-weighted dice loss was able to segment the notary sign sufficiently.

Vincent Christlein, Nikolaus Weichselbaumer, Saskia Limbach et al.

The type used to print an early modern book can give scholars valuable information about the time and place of its production as well as its producer. Recognizing such type is currently done manually using both the character shapes of `M' or `Qu' and the size of the total type to look it up in a large reference work. This is a reliable method, but it is also slow and requires specific skills. We investigate the performance of type classification and type retrieval using a newly created dataset consisting of easy and difficult types used in early printed books. For type classification, we rely on a deep Convolutional Neural Network (CNN) originally used for font-group classification while we use a common writer identification method for the retrieval case. We show that in both scenarios, easy types can be classified/retrieved with a high accuracy while difficult cases are indeed difficult.

Simon Jordan, Mathias Seuret, Pavel Král et al.

Automatic writer identification is a common problem in document analysis. State-of-the-art methods typically focus on the feature extraction step with traditional or deep-learning-based techniques. In retrieval problems, re-ranking is a commonly used technique to improve the results. Re-ranking refines an initial ranking result by using the knowledge contained in the ranked result, e. g., by exploiting nearest neighbor relations. To the best of our knowledge, re-ranking has not been used for writer identification/retrieval. A possible reason might be that publicly available benchmark datasets contain only few samples per writer which makes a re-ranking less promising. We show that a re-ranking step based on k-reciprocal nearest neighbor relationships is advantageous for writer identification, even if only a few samples per writer are available. We use these reciprocal relationships in two ways: encode them into new vectors, as originally proposed, or integrate them in terms of query-expansion. We show that both techniques outperform the baseline results in terms of mAP on three writer identification datasets.

Martin Mayr, Martin Stumpf, Anguelos Nicolaou et al.

Most people think that their handwriting is unique and cannot be imitated by machines, especially not using completely new content. Current cursive handwriting synthesis is visually limited or needs user interaction. We show that subdividing the process into smaller subtasks makes it possible to imitate someone's handwriting with a high chance to be visually indistinguishable for humans. Therefore, a given handwritten sample will be used as the target style. This sample is transferred to an online sequence. Then, a method for online handwriting synthesis is used to produce a new realistic-looking text primed with the online input sequence. This new text is then rendered and style-adapted to the input pen. We show the effectiveness of the pipeline by generating in- and out-of-vocabulary handwritten samples that are validated in a comprehensive user study. Additionally, we show that also a typical writer identification system can partially be fooled by the created fake handwritings.

Vincent Christlein, Anguelos Nicolaou, Mathias Seuret et al.

This competition investigates the performance of large-scale retrieval of historical document images based on writing style. Based on large image data sets provided by cultural heritage institutions and digital libraries, providing a total of 20 000 document images representing about 10 000 writers, divided in three types: writers of (i) manuscript books, (ii) letters, (iii) charters and legal documents. We focus on the task of automatic image retrieval to simulate common scenarios of humanities research, such as writer retrieval. The most teams submitted traditional methods not using deep learning techniques. The competition results show that a combination of methods is outperforming single methods. Furthermore, letters are much more difficult to retrieve than manuscripts.

Vincent Christlein, Lukas Spranger, Mathias Seuret et al.

Global pooling layers are an essential part of Convolutional Neural Networks (CNN). They are used to aggregate activations of spatial locations to produce a fixed-size vector in several state-of-the-art CNNs. Global average pooling or global max pooling are commonly used for converting convolutional features of variable size images to a fix-sized embedding. However, both pooling layer types are computed spatially independent: each individual activation map is pooled and thus activations of different locations are pooled together. In contrast, we propose Deep Generalized Max Pooling that balances the contribution of all activations of a spatially coherent region by re-weighting all descriptors so that the impact of frequent and rare ones is equalized. We show that this layer is superior to both average and max pooling on the classification of Latin medieval manuscripts (CLAMM'16, CLAMM'17), as well as writer identification (Historical-WI'17).

Michele Alberti, Lars Vögtlin, Vinaychandran Pondenkandath et al.

This paper introduces a new way for text-line extraction by integrating deep-learning based pre-classification and state-of-the-art segmentation methods. Text-line extraction in complex handwritten documents poses a significant challenge, even to the most modern computer vision algorithms. Historical manuscripts are a particularly hard class of documents as they present several forms of noise, such as degradation, bleed-through, interlinear glosses, and elaborated scripts. In this work, we propose a novel method which uses semantic segmentation at pixel level as intermediate task, followed by a text-line extraction step. We measured the performance of our method on a recent dataset of challenging medieval manuscripts and surpassed state-of-the-art results by reducing the error by 80.7%. Furthermore, we demonstrate the effectiveness of our approach on various other datasets written in different scripts. Hence, our contribution is two-fold. First, we demonstrate that semantic pixel segmentation can be used as strong denoising pre-processing step before performing text line extraction. Second, we introduce a novel, simple and robust algorithm that leverages the high-quality semantic segmentation to achieve a text-line extraction performance of 99.42% line IU on a challenging dataset.

Michele Alberti, Vinaychandran Pondenkandath, Marcel Würsch et al.

We propose a novel approach towards adversarial attacks on neural networks (NN), focusing on tampering the data used for training instead of generating attacks on trained models. Our network-agnostic method creates a backdoor during training which can be exploited at test time to force a neural network to exhibit abnormal behaviour. We demonstrate on two widely used datasets (CIFAR-10 and SVHN) that a universal modification of just one pixel per image for all the images of a class in the training set is enough to corrupt the training procedure of several state-of-the-art deep neural networks causing the networks to misclassify any images to which the modification is applied. Our aim is to bring to the attention of the machine learning community, the possibility that even learning-based methods that are personally trained on public datasets can be subject to attacks by a skillful adversary.

Michele Alberti, Mathias Seuret, Rolf Ingold et al.

The point of this paper is to question typical assumptions in deep learning and suggest alternatives. A particular contribution is to prove that even if a Stacked Convolutional Auto-Encoder is good at reconstructing pictures, it is not necessarily good at discriminating their classes. When using Auto-Encoders, intuitively one assumes that features which are good for reconstruction will also lead to high classification accuracy. Indeed, it became research practice and is a suggested strategy by introductory books. However, we prove that this is not always the case. We thoroughly investigate the quality of features produced by Stacked Convolutional Auto-Encoders when trained to reconstruct their input. In particular, we analyze the relation between the reconstruction and classification capabilities of the network, if we were to use the same features for both tasks. Experimental results suggest that in fact, there is no correlation between the reconstruction score and the quality of features for a classification task. This means, more formally, that the sub-dimension representation space learned from the Stacked Convolutional Auto-Encoder (while being trained for input reconstruction) is not necessarily better separable than the initial input space. Furthermore, we show that the reconstruction error is not a good metric to assess the quality of features, because it is biased by the decoder quality. We do not question the usefulness of pre-training, but we conclude that aiming for the lowest reconstruction error is not necessarily a good idea if afterwards one performs a classification task.

Michele Alberti, Mathias Seuret, Vinaychandran Pondenkandath et al.

In this paper, we present a novel approach to perform deep neural networks layer-wise weight initialization using Linear Discriminant Analysis (LDA). Typically, the weights of a deep neural network are initialized with: random values, greedy layer-wise pre-training (usually as Deep Belief Network or as auto-encoder) or by re-using the layers from another network (transfer learning). Hence, many training epochs are needed before meaningful weights are learned, or a rather similar dataset is required for seeding a fine-tuning of transfer learning. In this paper, we describe how to turn an LDA into either a neural layer or a classification layer. We analyze the initialization technique on historical documents. First, we show that an LDA-based initialization is quick and leads to a very stable initialization. Furthermore, for the task of layout analysis at pixel level, we investigate the effectiveness of LDA-based initialization and show that it outperforms state-of-the-art random weight initialization methods.

Kai Chen, Mathias Seuret

This paper presents a Convolutional Neural Network (CNN) based page segmentation method for handwritten historical document images. We consider page segmentation as a pixel labeling problem, i.e., each pixel is classified as one of the predefined classes. Traditional methods in this area rely on carefully hand-crafted features or large amounts of prior knowledge. In contrast, we propose to learn features from raw image pixels using a CNN. While many researchers focus on developing deep CNN architectures to solve different problems, we train a simple CNN with only one convolution layer. We show that the simple architecture achieves competitive results against other deep architectures on different public datasets. Experiments also demonstrate the effectiveness and superiority of the proposed method compared to previous methods.

Michele Alberti, Mathias Seuret, Rolf Ingold et al.

The point of this paper is to question typical assumptions in deep learning and suggest alternatives. A particular contribution is to prove that even if a Stacked Convolutional Auto-Encoder is good at reconstructing pictures, it is not necessarily good at discriminating their classes. When using Auto-Encoders, intuitively one assumes that features which are good for reconstruction will also lead to high classification accuracy. Indeed, it became research practice and is a suggested strategy by introductory books. However, we prove that this is not always the case. We thoroughly investigate the quality of features produced by Stacked Convolutional Auto-Encoders when trained to reconstruct their input. In particular, we analyze the relation between the reconstruction and classification capabilities of the network, if we were to use the same features for both tasks. Experimental results suggest that in fact, there is no correlation between the reconstruction score and the quality of features for a classification task. This means, more formally, that the sub-dimension representation space learned from the Stacked Convolutional Auto-Encoder (while being trained for input reconstruction) is not necessarily better separable than the initial input space. Furthermore, we show that the reconstruction error is not a good metric to assess the quality of features, because it is biased by the decoder quality. We do not question the usefulness of pre-training, but we conclude that aiming for the lowest reconstruction error is not necessarily a good idea if afterwards one performs a classification task.

Mathias Seuret, Michele Alberti, Rolf Ingold et al.

In this paper, we present a novel approach for initializing deep neural networks, i.e., by turning PCA into neural layers. Usually, the initialization of the weights of a deep neural network is done in one of the three following ways: 1) with random values, 2) layer-wise, usually as Deep Belief Network or as auto-encoder, and 3) re-use of layers from another network (transfer learning). Therefore, typically, many training epochs are needed before meaningful weights are learned, or a rather similar dataset is required for seeding a fine-tuning of transfer learning. In this paper, we describe how to turn a PCA into an auto-encoder, by generating an encoder layer of the PCA parameters and furthermore adding a decoding layer. We analyze the initialization technique on real documents. First, we show that a PCA-based initialization is quick and leads to a very stable initialization. Furthermore, for the task of layout analysis we investigate the effectiveness of PCA-based initialization and show that it outperforms state-of-the-art random weight initialization methods.