Abstract Text line segmentation in historical documents remains a significant challenge due to degraded manuscripts, complex layouts, and diverse handwriting styles. Developing robust computational methods is hindered by the scarcity of high-quality ground truth annotations, which require expert knowledge and are time-intensive to produce. Few-shot learning has emerged as a promising solution by enabling model training with minimal annotated data, yet its application to historical document analysis is still largely unexplored. To address this limitation, we introduce U-DIADS-TL (Uniud - Document Image Analysis DataSet - Text Line), a dataset specifically designed for text line segmentation in ancient manuscripts. U-DIADS-TL provides noise-free annotations with non-overlapping text elements and accommodates diverse document structures, including multi-column layouts. To encourage few-shot learning approaches, we offer only three training images, allowing researchers to develop segmentation models that can generalize from limited supervision. Our dataset serves as a critical bridge between deep learning and historical document analysis, fostering the creation of efficient, adaptable segmentation models for real-world applications.

Data efficiency remains a central challenge in brain anomaly detection, where annotated datasets are often scarce. Most existing methods are tailored to single-class settings and show limited ability to generalize. We introduce FsBAD, a feature reconstruction-based approach designed for few-shot brain anomaly detection with minimal supervision. FsBAD reconstructs a nominal version of an anomalous brain scan by leveraging a small set of aligned reference samples. To enhance reconstruction quality, we propose a novel feature alignment strategy that integrates regression with distribution regularization, promoting both semantic accuracy and nominal consistency. While FsBAD is optimized for brain imaging, we evaluate its generalization capabilities on liver and retina datasets. Experiments across all three domains show that FsBAD consistently outperforms state-of-the-art methods in both image-wise classification and pixel-wise anomaly localization, even in extremely low-shot (2- to 15-shot) settings. This demonstrates FsBAD’s potential as a scalable, data-efficient solution for brain anomaly detection and its robustness across medical imaging tasks.

Abstract Data availability is a big concern in the field of document analysis, especially when working on tasks that require a high degree of precision when it comes to the definition of the ground truths on which to train deep learning models. A notable example is represented by the task of document layout analysis in handwritten documents, which requires pixel-precise segmentation maps to highlight the different layout components of each document page. These segmentation maps are typically very time-consuming and require a high degree of domain knowledge to be defined, as they are intrinsically characterized by the content of the text. For this reason in the present work, we explore the effects of different initialization strategies for deep learning models employed for this type of task by relying on both in-domain and cross-domain datasets for their pre-training. To test the employed models we use two publicly available datasets with heterogeneous characteristics both regarding their structure as well as the languages of the contained documents. We show how a combination of cross-domain and in-domain transfer learning approaches leads to the best overall performance of the models, as well as speeding up their convergence process.

Accurate anomaly detection in brain Magnetic Resonance Imaging (MRI) is crucial for early diagnosis of neurological disorders, yet remains a significant challenge due to the high heterogeneity of brain abnormalities and the scarcity of annotated data. Traditional one-class classification models require extensive training on normal samples, limiting their adaptability to diverse clinical cases. In this work, we introduce MadIRC, an unsupervised anomaly detection framework that leverages Inter-Realization Channels (IRC) to construct a robust nominal model without any reliance on labeled data. We extensively evaluate MadIRC on brain MRI as the primary application domain, achieving a localization AUROC of 0.96 outperforming state-of-the-art supervised anomaly detection methods. Additionally, we further validate our approach on liver CT and retinal images to assess its generalizability across medical imaging modalities. Our results demonstrate that MadIRC provides a scalable, label-free solution for brain MRI anomaly detection, offering a promising avenue for integration into real-world clinical workflows.

Document layout segmentation is a challenging task due to the variability and complexity of document layouts. Ancient manuscripts in particular are often damaged by age, have very irregular layouts, and are characterized by progressive editing from different authors over a large time window. All these factors make the semantic segmentation process of specific areas, such as main text and side text, very difficult. However, the study of these manuscripts turns out to be fundamental for historians and humanists, so much so that in recent years the demand for machine learning approaches aimed at simplifying the extraction of information from these documents has consistently increased, leading document layout analysis to become an increasingly important research area. In order for machine learning techniques to be applied effectively to this task, however, a large amount of correctly and precisely labeled images is required for their training. This is obviously a limitation for this field of research as ground truth must be precisely and manually crafted by expert humanists, making it a very time-consuming process. In this paper, with the aim of overcoming this limitation, we present an efficient document layout segmentation framework, which while being trained on only one labeled page per manuscript still achieves state-of-the-art performance compared to other popular approaches trained on all the available data when tested on a challenging dataset of ancient Arabic manuscripts.

Layout analysis is a task of uttermost importance in ancient handwritten document analysis and represents a fundamental step toward the simplification of subsequent tasks such as optical character recognition and automatic transcription. However, many of the approaches adopted to solve this problem rely on a fully supervised learning paradigm. While these systems achieve very good performance on this task, the drawback is that pixel-precise text labeling of the entire training set is a very time-consuming process, which makes this type of information rarely available in a real-world scenario. In the present paper, we address this problem by proposing an efficient few-shot learning framework that achieves performances comparable to current state-of-the-art fully supervised methods on the publicly available DIVA-HisDB dataset.

Over the years, the humanities community has increasingly requested the creation of artificial intelligence frameworks to help the study of cultural heritage. Document Layout segmentation, which aims at identifying the different structural components of a document page, is a particularly interesting task connected to this trend, specifically when it comes to handwritten texts. While there are many effective approaches to this problem, they all rely on large amounts of data for the training of the underlying models, which is rarely possible in a real-world scenario, as the process of producing the ground truth segmentation task with the required precision to the pixel level is a very time-consuming task and often requires a certain degree of domain knowledge regarding the documents at hand. For this reason, in this paper, we propose an effective few-shot learning framework for document layout segmentation relying on two novel components, namely a dynamic instance generation and a segmentation refinement module. This approach is able of achieving performances comparable to the current state of the art on the popular Diva-HisDB dataset, while relying on just a fraction of the available data.

Image anomaly detection consists in detecting images or image portions that are visually different from the majority of the samples in a dataset. The task is of practical importance for various real-life applications like biomedical image analysis, visual inspection in industrial production, banking, traffic management, etc. Most of the current deep learning approaches rely on image reconstruction: the input image is projected in some latent space and then reconstructed, assuming that the network (mostly trained on normal data) will not be able to reconstruct the anomalous portions. However, this assumption does not always hold. We thus propose a new model based on the Vision Transformer architecture with patch masking: the input image is split in several patches, and each patch is reconstructed only from the surrounding data, thus ignoring the potentially anomalous information contained in the patch itself. We then show that multi-resolution patches and their collective embeddings provide a large improvement in the model's performance compared to the exclusive use of the traditional square patches. The proposed model has been tested on popular anomaly detection datasets such as MVTec and head CT and achieved good results when compared to other state-of-the-art approaches.