Michal Hradiš

Jan Kohút, Michal Hradiš

In many machine learning tasks, a large general dataset and a small specialized dataset are available. In such situations, various domain adaptation methods can be used to adapt a general model to the target dataset. We show that in the case of neural networks trained for handwriting recognition using CTC, simple fine-tuning with data augmentation works surprisingly well in such scenarios and that it is resistant to overfitting even for very small target domain datasets. We evaluated the behavior of fine-tuning with respect to augmentation, training data size, and quality of the pre-trained network, both in writer-dependent and writer-independent settings. On a large real-world dataset, fine-tuning on new writers provided an average relative CER improvement of 25 % for 16 text lines and 50 % for 256 text lines.

Jan Kohút, Michal Hradiš, Martin Kišš

One of the challenges of handwriting recognition is to transcribe a large number of vastly different writing styles. State-of-the-art approaches do not explicitly use information about the writer's style, which may be limiting overall accuracy due to various ambiguities. We explore models with writer-dependent parameters which take the writer's identity as an additional input. The proposed models can be trained on datasets with partitions likely written by a single author (e.g. single letter, diary, or chronicle). We propose a Writer Style Block (WSB), an adaptive instance normalization layer conditioned on learned embeddings of the partitions. We experimented with various placements and settings of WSB and contrastively pre-trained embeddings. We show that our approach outperforms a baseline with no WSB in a writer-dependent scenario and that it is possible to estimate embeddings for new writers. However, domain adaptation using simple fine-tuning in a writer-independent setting provides superior accuracy at a similar computational cost. The proposed approach should be further investigated in terms of training stability and embedding regularization to overcome such a baseline.

Martin Kišš, Michal Hradiš, Karel Beneš et al.

This paper explores semi-supervised training for sequence tasks, such as Optical Character Recognition or Automatic Speech Recognition. We propose a novel loss function $\unicode{x2013}$ SoftCTC $\unicode{x2013}$ which is an extension of CTC allowing to consider multiple transcription variants at the same time. This allows to omit the confidence based filtering step which is otherwise a crucial component of pseudo-labeling approaches to semi-supervised learning. We demonstrate the effectiveness of our method on a challenging handwriting recognition task and conclude that SoftCTC matches the performance of a finely-tuned filtering based pipeline. We also evaluated SoftCTC in terms of computational efficiency, concluding that it is significantly more efficient than a naïve CTC-based approach for training on multiple transcription variants, and we make our GPU implementation public.

Martin Kišš, Michal Hradiš

In this paper, we investigate self-supervised pre-training methods for document text recognition. Nowadays, large unlabeled datasets can be collected for many research tasks, including text recognition, but it is costly to annotate them. Therefore, methods utilizing unlabeled data are researched. We study self-supervised pre-training methods based on masked label prediction using three different approaches -- Feature Quantization, VQ-VAE, and Post-Quantized AE. We also investigate joint-embedding approaches with VICReg and NT-Xent objectives, for which we propose an image shifting technique to prevent model collapse where it relies solely on positional encoding while completely ignoring the input image. We perform our experiments on historical handwritten (Bentham) and historical printed datasets mainly to investigate the benefits of the self-supervised pre-training techniques with different amounts of annotated target domain data. We use transfer learning as strong baselines. The evaluation shows that the self-supervised pre-training on data from the target domain is very effective, but it struggles to outperform transfer learning from closely related domains. This paper is one of the first researches exploring self-supervised pre-training in document text recognition, and we believe that it will become a cornerstone for future research in this area. We made our implementation of the investigated methods publicly available at https://github.com/DCGM/pero-pretraining.

Martin Kostelník, Michal Hradiš, Martin Dočekal

Topic localization aims to identify spans of text that express a given topic defined by a name and description. To study this task, we introduce a human-annotated benchmark based on Czech historical documents, containing human-defined topics together with manually annotated spans and supporting evaluation at both document and word levels. Evaluation is performed relative to human agreement rather than a single reference annotation. We evaluate a diverse range of large language models alongside BERT-based models fine-tuned on a distilled development dataset. Results reveal substantial variability among LLMs, with performance ranging from near-human topic detection to pronounced failures in span localization. While the strongest models approach human agreement, the distilled token embedding models remain competitive despite their smaller scale. The dataset and evaluation framework are publicly available at: https://github.com/dcgm/czechtopic.

Martin Kostelník, Karel Beneš, Michal Hradiš

Logical page segmentation is an important step in document analysis, enabling better semantic representations, information retrieval, and text understanding. Previous approaches define logical segmentation either through text or geometric objects, relying on OCR or precise geometry. To avoid the need for OCR, we define the task purely as segmentation in the image domain. Furthermore, to ensure the evaluation remains unaffected by geometrical variations that do not impact text segmentation, we propose to use only foreground text pixels in the evaluation metric and disregard all background pixels. To support research in logical document segmentation, we introduce TextBite, a dataset of historical Czech documents spanning the 18th to 20th centuries, featuring diverse layouts from newspapers, dictionaries, and handwritten records. The dataset comprises 8,449 page images with 78,863 annotated segments of logically and thematically coherent text. We propose a set of baseline methods combining text region detection and relation prediction. The dataset, baselines and evaluation framework can be accessed at https://github.com/DCGM/textbite-dataset.

Jan Kohút, Martin Dočekal, Michal Hradiš et al.

Manual digitization of bibliographic metadata is time consuming and labor intensive, especially for historical and real-world archives with highly variable formatting across documents. Despite advances in machine learning, the absence of dedicated datasets for metadata extraction hinders automation. To address this gap, we introduce BiblioPage, a dataset of scanned title pages annotated with structured bibliographic metadata. The dataset consists of approximately 2,000 monograph title pages collected from 14 Czech libraries, spanning a wide range of publication periods, typographic styles, and layout structures. Each title page is annotated with 16 bibliographic attributes, including title, contributors, and publication metadata, along with precise positional information in the form of bounding boxes. To extract structured information from this dataset, we valuated object detection models such as YOLO and DETR combined with transformer-based OCR, achieving a maximum mAP of 52 and an F1 score of 59. Additionally, we assess the performance of various visual large language models, including LlamA 3.2-Vision and GPT-4o, with the best model reaching an F1 score of 67. BiblioPage serves as a real-world benchmark for bibliographic metadata extraction, contributing to document understanding, document question answering, and document information extraction. Dataset and evaluation scripts are availible at: https://github.com/DCGM/biblio-dataset

Martin Kišš, Michal Hradiš, Martina Dvořáková et al.

We introduce the AnnoPage Dataset, a novel collection of 7,550 pages from historical documents, primarily in Czech and German, spanning from 1485 to the present, focusing on the late 19th and early 20th centuries. The dataset is designed to support research in document layout analysis and object detection. Each page is annotated with axis-aligned bounding boxes (AABB) representing elements of 25 categories of non-textual elements, such as images, maps, decorative elements, or charts, following the Czech Methodology of image document processing. The annotations were created by expert librarians to ensure accuracy and consistency. The dataset also incorporates pages from multiple, mainly historical, document datasets to enhance variability and maintain continuity. The dataset is divided into development and test subsets, with the test set carefully selected to maintain the category distribution. We provide baseline results using YOLO and DETR object detectors, offering a reference point for future research. The AnnoPage Dataset is publicly available on Zenodo (https://doi.org/10.5281/zenodo.12788419), along with ground-truth annotations in YOLO format.

Marek Vaško, Adam Herout, Michal Hradiš

When digitizing historical archives, it is necessary to search for the faces of celebrities and ordinary people, especially in newspapers, link them to the surrounding text, and make them searchable. Existing face detectors on datasets of scanned historical documents fail remarkably -- current detection tools only achieve around 24% mAP at 50:90% IoU. This work compensates for this failure by introducing a new manually annotated domain-specific dataset in the style of the popular Wider Face dataset, containing 2.2k new images from digitized historical newspapers from the 19th to 20th century, with 11k new bounding-box annotations and associated facial landmarks. This dataset allows existing detectors to be retrained to bring their results closer to the standard in the field of face detection in the wild. We report several experimental results comparing different families of fine-tuned detectors against publicly available pre-trained face detectors and ablation studies of multiple detector sizes with comprehensive detection and landmark prediction performance results.

Martin Kišš, Michal Hradiš

Self-supervised learning has emerged as a powerful approach for leveraging large-scale unlabeled data to improve model performance in various domains. In this paper, we explore masked self-supervised pre-training for text recognition transformers. Specifically, we propose two modifications to the pre-training phase: progressively increasing the masking probability, and modifying the loss function to incorporate both masked and non-masked patches. We conduct extensive experiments using a dataset of 50M unlabeled text lines for pre-training and four differently sized annotated datasets for fine-tuning. Furthermore, we compare our pre-trained models against those trained with transfer learning, demonstrating the effectiveness of the self-supervised pre-training. In particular, pre-training consistently improves the character error rate of models, in some cases up to 30 % relatively. It is also on par with transfer learning but without relying on extra annotated text lines.

Jan Kohút, Michal Hradiš

A common use case for OCR applications involves users uploading documents and progressively correcting automatic recognition to obtain the final transcript. This correction phase presents an opportunity for progressive adaptation of the OCR model, making it crucial to adapt early, while ensuring stability and reliability. We demonstrate that state-of-the-art transformer-based models can effectively support this adaptation, gradually reducing the annotator's workload. Our results show that fine-tuning can reliably start with just 16 lines, yielding a 10% relative improvement in CER, and scale up to 40% with 256 lines. We further investigate the impact of model components, clarifying the roles of the encoder and decoder in the fine-tuning process. To guide adaptation, we propose reliable stopping criteria, considering both direct approaches and global trend analysis. Additionally, we show that OCR models can be leveraged to cut annotation costs by half through confidence-based selection of informative lines, achieving the same performance with fewer annotations.

Martin Kišš, Jan Kohút, Karel Beneš et al.

This paper describes a system prepared at Brno University of Technology for ICDAR 2021 Competition on Historical Document Classification, experiments leading to its design, and the main findings. The solved tasks include script and font classification, document origin localization, and dating. We combined patch-level and line-level approaches, where the line-level system utilizes an existing, publicly available page layout analysis engine. In both systems, neural networks provide local predictions which are combined into page-level decisions, and the results of both systems are fused using linear or log-linear interpolation. We propose loss functions suitable for weakly supervised classification problem where multiple possible labels are provided, and we propose loss functions suitable for interval regression in the dating task. The line-level system significantly improves results in script and font classification and in the dating task. The full system achieved 98.48 %, 88.84 %, and 79.69 % accuracy in the font, script, and location classification tasks respectively. In the dating task, our system achieved a mean absolute error of 21.91 years. Our system achieved the best results in all tasks and became the overall winner of the competition.

Martin Kišš, Karel Beneš, Michal Hradiš

This paper addresses text recognition for domains with limited manual annotations by a simple self-training strategy. Our approach should reduce human annotation effort when target domain data is plentiful, such as when transcribing a collection of single person's correspondence or a large manuscript. We propose to train a seed system on large scale data from related domains mixed with available annotated data from the target domain. The seed system transcribes the unannotated data from the target domain which is then used to train a better system. We study several confidence measures and eventually decide to use the posterior probability of a transcription for data selection. Additionally, we propose to augment the data using an aggressive masking scheme. By self-training, we achieve up to 55 % reduction in character error rate for handwritten data and up to 38 % on printed data. The masking augmentation itself reduces the error rate by about 10 % and its effect is better pronounced in case of difficult handwritten data.

Jan Kohút, Michal Hradiš

Users of OCR systems, from different institutions and scientific disciplines, prefer and produce different transcription styles. This presents a problem for training of consistent text recognition neural networks on real-world data. We propose to extend existing text recognition networks with a Transcription Style Block (TSB) which can learn from data to switch between multiple transcription styles without any explicit knowledge of transcription rules. TSB is an adaptive instance normalization conditioned by identifiers representing consistently transcribed documents (e.g. single document, documents by a single transcriber, or an institution). We show that TSB is able to learn completely different transcription styles in controlled experiments on artificial data, it improves text recognition accuracy on large-scale real-world data, and it learns semantically meaningful transcription style embedding. We also show how TSB can efficiently adapt to transcription styles of new documents from transcriptions of only a few text lines.

Oldřich Kodym, Michal Hradiš

Extraction of text regions and individual text lines from historic documents is necessary for automatic transcription. We propose extending a CNN-based text baseline detection system by adding line height and text block boundary predictions to the model output, allowing the system to extract more comprehensive layout information. We also show that pixel-wise text orientation prediction can be used for processing documents with multiple text orientations. We demonstrate that the proposed method performs well on the cBAD baseline detection dataset. Additionally, we benchmark the method on newly introduced PERO layout dataset which we also make public.

Martin Kišš, Michal Hradiš, Oldřich Kodym

We introduce the Brno Mobile OCR Dataset (B-MOD) for document Optical Character Recognition from low-quality images captured by handheld mobile devices. While OCR of high-quality scanned documents is a mature field where many commercial tools are available, and large datasets of text in the wild exist, no existing datasets can be used to develop and test document OCR methods robust to non-uniform lighting, image blur, strong noise, built-in denoising, sharpening, compression and other artifacts present in many photographs from mobile devices. This dataset contains 2 113 unique pages from random scientific papers, which were photographed by multiple people using 23 different mobile devices. The resulting 19 728 photographs of various visual quality are accompanied by precise positions and text annotations of 500k text lines. We further provide an evaluation methodology, including an evaluation server and a testset with non-public annotations. We provide a state-of-the-art text recognition baseline build on convolutional and recurrent neural networks trained with Connectionist Temporal Classification loss. This baseline achieves 2 %, 22 % and 73 % word error rates on easy, medium and hard parts of the dataset, respectively, confirming that the dataset is challenging. The presented dataset will enable future development and evaluation of document analysis for low-quality images. It is primarily intended for line-level text recognition, and can be further used for line localization, layout analysis, image restoration and text binarization.

Martin Kolář, Michal Hradiš, Pavel Zemčík

This report presents our submission to the MS COCO Captioning Challenge 2015. The method uses Convolutional Neural Network activations as an embedding to find semantically similar images. From these images, the most typical caption is selected based on unigram frequencies. Although the method received low scores with automated evaluation metrics and in human assessed average correctness, it is competitive in the ratio of captions which pass the Turing test and which are assessed as better or equal to human captions.