Sign language translation systems traditionally rely on intermediate gloss representations to bridge the gap between visual input and written language output. However, manual gloss annotation is costly, language-dependent, and often lossy, prompting growing interest in gloss-free alternatives. This paper introduces H andscribe , a novel two-stage framework for gloss-free sign language translation and gloss sequence generation. H andscribe first translates continuous sign language videos into written language sentences using a lightweight decoder built atop SlowFast-based spatiotemporal features and a frozen mBART model. Then, in the second stage, it generates gloss sequences from these sentences using a Large Language Model (LLaMa3.1-8B-Instruct) that has been fine-tuned with weak supervision. Our experiments on PHOENIX-2014-T and Wav2Gloss Fieldwork demonstrate strong translation performance and state-of-the-art multilingual gloss generation, even in zero-shot settings. The proposed framework reduces annotation bottlenecks while maintaining flexibility and interpretability, paving the way for scalable and inclusive sign language technologies. The code and fine-tuning scripts are available at https://github.com/colonnaemanuele/Handscribe .

Abstract Balancing predictive performance with transparency remains a core challenge in eXplainable Artificial Intelligence, especially for tabular data. In this work, we present FuzzyGNN, a novel hybrid framework that combines fuzzy logic with Graph Neural Networks (GNNs) to deliver both high accuracy and interpretability. FuzzyGNN represents each data instance as a graph, where raw features and fuzzy sets are modeled as nodes, enabling contextual reasoning through message passing. Unlike traditional neuro-fuzzy systems—often limited by scalability and rule explosion—FuzzyGNN enforces model compactness through its graph-aware architecture and a pruning mechanism guided by graphbased explanations. Extensive experiments on seven benchmark datasets show that FuzzyGNN achieves competitive predictive performance while drastically reducing the number of fuzzy rules, linguistic terms, and active features used to produce output fuzzy rules. Additionally, it offers structural interpretability through rule extraction and post-hoc explainability via GNNExplainer. These capabilities make FuzzyGNN particularly well-suited for domains where transparency is as crucial as accuracy, such as healthcare, finance, and safety-critical systems. The code is available in the following link: https://github.com/cilabuniba/fuzzygnn

Few-shot semantic segmentation (FSS) aims to segment novel classes in query images using only a small annotated support set. While prior research has mainly focused on improving decoders, the encoder’s limited ability to extract meaningful features for unseen classes remains a key bottleneck. In this work, we introduce Take a Peek (TaP), a simple yet effective method that enhances encoder adaptability for both FSS and cross-domain FSS by inducing a lightweight feature-space shift conditioned on the support set. TaP leverages Low-Rank Adaptation to fine-tune the encoder on the support set with minimal computational overhead, enabling fast adaptation to novel classes while mitigating catastrophic forgetting. Our method is model-agnostic and can be seamlessly integrated into existing FSS pipelines. Extensive experiments across multiple benchmarks-including COCO 20 i , Pascal 5 i , and cross-domain datasets such as DeepGlobe, ISIC, and Chest X-ray-demonstrate that TaP consistently improves segmentation performance across diverse models and shot settings. Notably, TaP delivers significant gains in complex multi-class scenarios, highlighting its practical effectiveness in realistic settings. A rank sensitivity analysis also shows that strong performance can be achieved even with low-rank adaptations, thereby ensuring computational efficiency. By addressing a critical limitation in FSS-the encoder’s generalization to novel classes-TaP paves the way toward more robust, efficient, and generalizable segmentation systems. The code is available at https://github.com/pasqualedem/TakeAPeek .

Few-shot semantic segmentation aims to segment objects from previously unseen classes using only a limited number of labeled examples. In this paper, we introduce Label Anything, a novel transformer-based architecture designed for multi-prompt, multi-way few-shot semantic segmentation. Our approach leverages diverse visual prompts—points, bounding boxes, and masks—to create a highly flexible and generalizable framework that significantly reduces annotation burden while maintaining high accuracy. Label Anything makes three key contributions: (i) we introduce a new task formulation that relaxes conventional few-shot segmentation constraints by supporting various types of prompts, multi-class classification, and enabling multiple prompts within a single image; (ii) we propose a novel architecture based on transformers and attention mechanisms; and (iii) we design a versatile training procedure allowing our model to operate seamlessly across different N-way K-shot and prompt-type configurations with a single trained model. Our extensive experimental evaluation on the widely used COCO-20i benchmark demonstrates that Label Anything achieves state-of-the-art performance among existing multi-way few-shot segmentation methods, while significantly outperforming leading single-class models when evaluated in multi-class settings. Code and trained models are available at https://github.com/pasqualedem/LabelAnything.

This work explores integrating large language models (LLMs) with the adaptive-network-based fuzzy inference system (ANFIS) to enhance interpretability and usability in decision-making processes. ANFIS generates transparent and interpretable fuzzy rules, while LLMs complement these by providing concise, context-aware textual explanations. By combining ANFIS’s data-driven capabilities with the semantic understanding of LLMs, the framework aims to clarify AI outputs and support error identification in the knowledge base. The pipeline implements a human-in-the-loop strategy to engage domain experts in enhancing prompts, verifying explanations, and aligning outputs with expert standards. The methodology was assessed in a medical setting, particularly for predicting epilepsy seizures using EEG data. This study illustrates how the proposed pipeline bridges AI models and real-world applications, providing transparent insights into decision-making processes. It lays the groundwork for creating more interactive, accurate, explainable, and user-friendly tools for predictive analytics, particularly in critical fields like healthcare.

Inpainting focuses on filling missing or corrupted regions of an image to blend seamlessly with its surrounding content and style. While conditional diffusion models have proven effective for text-guided inpainting, we introduce the novel task of multi-mask inpainting, where multiple regions are simultaneously inpainted using distinct prompts. Furthermore, we design a fine-tuning procedure for multimodal LLMs, such as LLaVA, to generate multi-mask prompts automatically using corrupted images as inputs. These models can generate helpful and detailed prompt suggestions for filling the masked regions. The generated prompts are then fed to Stable Diffusion, which is fine-tuned for the multi-mask inpainting problem using rectified cross-attention, enforcing prompts onto their designated regions for filling. Experiments on digitized paintings from WikiArt and the Densely Captioned Images dataset demonstrate that our pipeline delivers creative and accurate inpainting results. Our code, data, and trained models are available at https://cilabuniba.github.io/i-dream-my-painting.

Abstract Crowdfunding has evolved into a formidable mechanism for collective financing, challenging traditional funding sources such as bank loans, venture capital, and private equity with its global reach and versatile applications across various sectors. This paper explores the complex dynamics of crowdfunding platforms, particularly focusing on investor behaviour and investment patterns within equity and lending campaigns in Italy. By leveraging advanced machine learning techniques, including XGBoost and LSTM networks, we develop predictive models that dynamically analyze real-time and historical data to accurately forecast the success or failure of crowdfunding campaigns. To address the existing gaps in crowdfunding analysis tools, we introduce two novel datasets—one for equity crowdfunding and another for lending. Moreover, our approach extends beyond traditional binary success metrics, proposing novel measures. The insights gained from this study could support crowdfunding strategies, significantly improving project selection and promotional tactics on platforms. By enhancing decision-making processes and providing forward-looking guidance to investors, our computational model aims to empower both campaign creators and platform administrators, ultimately improving the overall efficacy and sustainability of crowdfunding as a financing tool.

We present GraphCLIP, a novel contrastive learning framework for multimodal artwork classification that integrates visual and contextual information to improve predictive accuracy and interpretability. Traditional computer vision methods often fall short in visual arts, where context is crucial. GraphCLIP leverages image data and a Knowledge Graph to extract features from both perspectives. Evaluated on the A r t G r a p h dataset, with over 100,000 artworks in 32 styles and 18 genres, GraphCLIP outperforms existing models in single-task (up to + 8 % in F1-score) and multi-task settings (up to + 6 % ), demonstrating robustness even with unseen classes. Additionally, visual and contextual qualitative explanations enhance model transparency. The versatility of GraphCLIP extends beyond art classification: its methodology can be adapted to other domains where integrating diverse data types is essential. (The code is publicly available at: https://github.com/CILAB-ArtGraph/graphclip.git .) • We introduce GraphCLIP, a contrastive learning framework for artwork classification. • GraphCLIP combines visual data with contextual knowledge. • We achieve state-of-the-art performance on the A r t G r a p h dataset. • We demonstrate robustness with unseen classes in distribution shift scenarios. • We provide visual and contextual explanations to enhance model interpretability.

Online signature parameters, which are based on human characteristics, broaden the applicability of an automatic signature verifier. Although kinematic and dynamic features have previously been suggested, accurately measuring features such as arm and forearm torques remains challenging. We present two approaches for estimating angular velocities, angular positions, and force torques. The first approach involves using a physical UR5e robotic arm to reproduce a signature while capturing those parameters over time. The second method, a cost-effective approach, uses a neural network to estimate the same parameters. Our findings demonstrate that a simple neural network model can extract effective parameters for signature verification. Training the neural network with the MCYT300 dataset and cross-validating with other databases, namely, BiosecurID, Visual, Blind, OnOffSigDevanagari-75 and OnOffSigBengali-75 confirm the model’s generalization capability. The trained model is available at: https://github.com/gvessio/SignatureKinematics . • We explore kinematic and dynamic features for online signature verification. • A UR5 robotic arm is used to acquire these features from the MCYT330 dataset. • A neural network estimates the kinematic and dynamic features of a signature. • We demonstrate promising performance using the estimated features across datasets.

Accurate brain tumor segmentation is crucial for precise medical diagnosis and treatment planning in medical imaging. This research delves into assessing the effectiveness and transparency of Graph Neural Networks (GNNs) in brain tumor segmentation. The primary objectives include comparing various GNN architectures and improving their understandability by applying the GNNExplainer method. Leveraging the BraTS 2021 challenge dataset, which consists of MRI scans and corresponding ground truth annotations, the study reveals the successful application of GNNs in achieving precise brain tumor segmentation. By incorporating GNNExplainer, the explainability of the models is significantly enhanced, shedding light on the decision-making processes within the network. The proposed approach could advance the field of brain tumor segmentation, providing clinicians with accurate and transparent models to inform their decision-making processes in patient care.1

Abstract Signature verification is a critical task in many applications, including forensic science, legal judgments, and financial markets. However, current signature verification systems are often difficult to explain, which can limit their acceptance in these applications. In this paper, we propose a novel explainable offline automatic signature verifier (ASV) to support forensic handwriting examiners. Our ASV is based on a universal background model (UBM) constructed from offline signature images. It allows us to assign a questioned signature to the UBM and to a reference set of known signatures using simple distance measures. This makes it possible to explain the verifier’s decision in a way that is understandable to non-experts. We evaluated our ASV on publicly available databases and found that it achieves competitive performance with state-of-the-art ASVs, even when challenging 1 versus 1 comparisons are considered. Our results demonstrate that it is possible to develop an explainable ASV that is also competitive in terms of performance. We believe that our ASV has the potential to improve the acceptance of signature verification in critical applications such as forensic science and legal judgments.

We propose a new approach for Alzheimer's disease (AD) detection using diffusion tensor imaging, specifically fractional anisotropy (FA) images, based on a combination of unsupervised and supervised deep learning techniques. Our method involves training a 3D convolutional autoencoder to learn low-dimensional representations of FA images in an unsupervised manner and using the learned representations to pre-train a supervised 3D convolutional classifier to predict the presence or absence of AD. Unsupervised pre-training can improve the classifier's performance, especially when difficult-to-collect labeled data are limited. We evaluate our approach on the OASIS-3 dataset and demonstrate promising performance.

In precision agriculture, non-invasive remote sensing using UAVs can be employed to observe crops in visible and nonvisible spectra.This paper investigates the effectiveness of stateof-the-art knowledge distillation techniques for mapping weeds with drones, an essential component of precision agriculture that employs remote sensing to monitor crops and weeds.The study introduces a lightweight Vision Transformer-based model that achieves optimal weed mapping capabilities while maintaining minimal computation time.The research shows that the student model effectively learns from the teacher model using the WeedMap dataset, achieving accurate results suitable for mobile platforms such as drones, with only 0.5 GMacs compared to 42.5 GMacs of the teacher model.The trained models obtained an F1 score of 0.863 and 0.631 on two data subsets, with a performance improvement of 2 and 7 points, respectively, over the undistilled model.The study results suggest that developing efficient computer vision algorithms on drones can significantly improve agricultural management practices, leading to greater profitability and environmental sustainability.

In precision agriculture, noninvasive remote sensing can be used to observe crops and weeds in visible and non-visible spectra. This paper proposes a novel approach for weed mapping using lightweight Vision Transformers. The method uses a lightweight Transformer architecture to process high-resolution aerial images obtained from drones and performs semantic segmentation to distinguish between crops and weeds. The method also employs specific architectural designs to enable transfer learning from RGB weights in a multispectral setting. For this purpose, the WeedMap dataset, acquired by drones equipped with multispectral cameras, was used. The experimental results demonstrate the effectiveness of the proposed method, exceeding the state-of-the-art. Our approach also enables more efficient mapping, allowing farmers to quickly and easily identify infested areas and prioritize their control efforts. These results encourage using drones as versatile computer vision flying devices for herbicide management, thereby improving crop yields. The code is available at https://github.com/pasqualedem/LWViTs-for-weedmapping.

In medical imaging, images represent the quantification of the interaction between electromagnetic waves and our body and are represented in grey-scale. In addition, medical imaging often produces multimodal images. However, the analysis and interpretation of these images mostly occur in sequence or, as in the case of automatic tools, they are simply concatenated as independent sources of information. In both cases, color perception and color contrast are not exploited. Color perception and color contrast play a crucial role in human vision to recognize objects effectively and efficiently, and this can in principle extend to automatic systems. In this paper we show how color coding, particularly using color opponent models, can become an effective tool for preliminary color-based segmentation. Tests have been conducted on multimodal Magnetic Resonance Imaging (MRI) of the brain collected in a public database and the results obtained show the importance of color coding in medical imaging analysis.

Crowd analysis from drones has attracted increasing attention in recent times, thanks to the ease of deployment and affordable cost of these devices. However, how this technology can provide a solution to crowd flow detection is still an explored research question. In this paper, we contribute by proposing a crowd flow detection method for video sequences shot by a drone. The method is mainly based on a Fully Convolutional Network model for crowd density estimation, which aims to provide a good compromise between effectiveness and efficiency, and clustering algorithms aimed at detecting the centroids of high-density areas in density maps. The method was tested on the VisDrone Crowd Counting dataset-characterized not by still images but by video sequences-providing promising results. This direction may open up new ways of analyzing high-level crowd behavior from drones.1

Abstract Clustering artworks is difficult for several reasons. On the one hand, recognizing meaningful patterns based on domain knowledge and visual perception is extremely hard. On the other hand, applying traditional clustering and feature reduction techniques to the highly dimensional pixel space can be ineffective. To address these issues, in this paper we propose : a DEep learning approach to cLustering vIsUal artS. The method uses a pre-trained convolutional network to extract features and then feeds these features into a deep embedded clustering model, where the task of mapping the input data to a latent space is jointly optimized with the task of finding a set of cluster centroids in this latent space. Quantitative and qualitative experimental results show the effectiveness of the proposed method. can be useful for several tasks related to art analysis, in particular visual link retrieval and historical knowledge discovery in painting datasets.

We introduce an approach called PLENARY (exPlaining bLack-box modEls in Natural lAnguage thRough fuzzY linguistic summaries), which is an explainable classifier based on a data-driven predictive model. Neural learning is exploited to derive a predictive model based on two levels of labels associated with the data. Then, model explanations are derived through the popular SHapley Additive exPlanations (SHAP) tool and conveyed in a linguistic form via fuzzy linguistic summaries. The linguistic summarization allows translating the explanations of the model outputs provided by SHAP into statements expressed in natural language. PLENARY accounts for the imprecision related to model outputs by summarizing them into simple linguistic statements and for the imprecision related to the data labeling process by including additional domain knowledge in the form of middle-layer labels. PLENARY is validated on preprocessed speech signals collected from smartphones from patients with bipolar disorder and on publicly available mental health survey data. The experiments confirm that fuzzy linguistic summarization is an effective technique to support meta-analyses of the outputs of AI models. Also, PLENARY improves explainability by aggregating low-level attributes into high-level information granules, and by incorporating vague domain knowledge into a multi-task sequential and compositional multilayer perceptron. SHAP explanations translated into fuzzy linguistic summaries significantly improve understanding of the predictive modelling process and its outputs.

The growing availability of large collections of digitized artworks has disclosed new opportunities to develop intelligent systems for the automatic analysis of fine arts. Among other benefits, these tools can foster a deeper understanding of fine arts, ultimately supporting the spread of culture. However, most of the systems proposed in the literature are only based on visual features of digitized artwork images, which are sometimes only integrated with some metadata and textual comments. A Knowledge Graph (KG) that integrates a rich body of information about artworks, artists, painting schools, etc., in a unified structured framework, can provide a valuable resource for more powerful information retrieval and knowledge discovery tools in the artistic domain. To this end, in this paper we present ArtGraph:1 an artistic KG based on WikiArt and DBpedia. The graph already provides knowledge discovery capabilities without having to train a learning system. In addition, we propose a novel KG-enabled fine art classification method based on ArtGraph, which is used to perform artwork attribute prediction tasks. The method extracts embeddings from ArtGraph and injects them as “contextual” knowledge into a Deep Learning model. Compared to the state-of-the-art, the proposed model provides encouraging results, suggesting that the exploitation of KGs in combination with Deep Learning can pave the way for bridging the gap between the Humanities and Computer Science communities.

Abstract To date, Artificial Intelligence systems for handwriting and drawing analysis have primarily targeted domains such as writer identification and sketch recognition. Conversely, the automatic characterization of graphomotor patterns asbiomarkersof brain health is a relatively less explored research area. Despite its importance, the work done in this direction is limited and sporadic. This paper aims to provide a survey of related work to provide guidance to novice researchers and highlight relevant study contributions. The literature has been grouped into “visual analysis techniques” and “procedural analysis techniques”. Visual analysis techniques evaluate offline samples of a graphomotor response after completion. On the other hand, procedural analysis techniques focus on the dynamic processes involved in producing a graphomotor reaction. Since the primary goal of both families of strategies is to represent domain knowledge effectively, the paper also outlines the commonly employed handwriting representation and estimation methods presented in the literature and discusses their strengths and weaknesses. It also highlights the limitations of existing processes and the challenges commonly faced when designing such systems. High-level directions for further research conclude the paper.

Abstract Visual arts are of inestimable importance for the cultural, historic and economic growth of our society. One of the building blocks of most analysis in visual arts is to find similarity relationships among paintings of different artists and painting schools. To help art historians better understand visual arts, this paper presents a framework for visual link retrieval and knowledge discovery in digital painting datasets. Visual link retrieval is accomplished by using a deep convolutional neural network to perform feature extraction and a fully unsupervised nearest neighbor mechanism to retrieve links among digitized paintings. Historical knowledge discovery is achieved by performing a graph analysis that makes it possible to study influences among artists. An experimental evaluation on a database collecting paintings by very popular artists shows the effectiveness of the method. The unsupervised strategy makes the method interesting especially in cases where metadata are scarce, unavailable or difficult to collect.

Abstract This paper provides an overview of some of the most relevant deep learning approaches to pattern extraction and recognition in visual arts, particularly painting and drawing. Recent advances in deep learning and computer vision, coupled with the growing availability of large digitized visual art collections, have opened new opportunities for computer science researchers to assist the art community with automatic tools to analyse and further understand visual arts. Among other benefits, a deeper understanding of visual arts has the potential to make them more accessible to a wider population, ultimately supporting the spread of culture.

Clustering artworks is difficult for several reasons. On the one hand, recognizing meaningful patterns in accordance with domain knowledge and visual perception is extremely difficult. On the other hand, applying traditional clustering and feature reduction techniques to the highly dimensional pixel space can be ineffective. To address these issues, we propose to use a deep convolutional embedding model for digitized painting clustering, in which the task of mapping the raw input data to an abstract, latent space is jointly optimized with the task of finding a set of cluster centroids in this latent feature space. Quantitative and qualitative experimental results show the effectiveness of the proposed method. The model is also capable of outperforming other state-of-the-art deep clustering approaches to the same problem. The proposed method can be useful for several art-related tasks, in particular visual link retrieval and historical knowledge discovery in painting datasets.

Crowd analysis is receiving an increasing attention in the last years because of its social and public safety implications. One of the building blocks of crowd analysis is crowd counting and the associated crowd density estimation. Several commercially available drones are equipped with onboard cameras and embed powerful GPUs, making them an excellent platform for real-time crowd counting tools. This paper proposes a light-weight and fast fully-convolutional neural network to learn a regression model for crowd counting in images acquired from drones. A robust model is derived by training the network from scratch on a subset of the very challenging VisDrone dataset, which is characterized by a high variety of locations, environments, perspectives and lighting conditions. The derived model achieves an MAE of 8.86 and an RMSE of 15.07 on the test images, outperforming models developed by state-of-the-art light-weight architectures, that are MobileNetV2 and YOLOv3.

Humans use a wide range of non-verbal social signals while communicating with each other. Gestures are part of these signals and social robots should be able to recognize them for responding appropriately during a dialogue and being more socially believable. Gesture recognition is a hot topic in Computer Vision since a long time. This is particularly due to the fact that the segmentation of foreground objects from a cluttered background is a challenging problem, especially if it has to be performed in real-time. In this paper, we propose a vision-based framework for making social robots capable of recognizing and responding in real-time to a specific greeting gesture, namely the hand waving. The framework is based on a Convolutional Neural Network model trained to recognize hand gestures. Preliminary experiments in a lab setup with the social robot Pepper indicate that the robot correctly recognizes the wave gesture 90% of the times and answers appropriately in real-time by waving itself, thus increasing its social believability.

Education may play a key role in developing “cognitive reserve” against neurodegenerative dementia. In this work, we investigate for the first time if handwriting dynamics can serve as a quantitative indicator of this reserve. We carried out an exploratory study involving a sample of mild cognitive impairment (MCI) subjects, with high and low education respectively, and a sample of healthy elder controls. We asked them to perform three complex handwriting tasks on a digitizing tablet: drawing a clock; copying a check; writing a spontaneous sentence. Dynamic measures of the handwriting were then analyzed both with an unsupervised and a supervised machine learning approach. The results we obtained suggest that: (i) handwriting of MCI subjects with high reserve is quite similar to that of controls; (ii) handwriting of MCI subjects with lower reserve is easier to be distinguished from the other two. Dynamic handwriting analysis could provide a novel methodology to elucidate the still unknown mechanisms underlying brain resilience.

In the present paper, we propose an evolutionary approach to address interoperability issues in multi-device signature verification, based on transformation mappings automatically tuned by a genetic algorithm. These mappings are meant to decrease dissimilarities between signatures acquired through different devices and with different modalities (stylus/finger). The effectiveness of the proposed method was evaluated on the e-BioSign data set. Our proposal achieved an average relative improvement of 26% of EER, for the case of skilled forgeries, compared to baseline results.

Mobile Ad-hoc NETworks (MANETs) are wireless networks designed for communications among nomadic hosts in absence of fixed infrastructure. Different reactive protocols for MANETs, in which routes are established only when needed, provide different network topology awareness (NTA) to each host, depending on their algorithmic features. NACK-based AODV (N-AODV) is a variant of the well-known Ad-hoc On-demand Distance Vector (AODV) reactive protocol for MANETs we proposed with the aim of improving NTA of the original protocol. In this paper, a performance comparison between AODV and N-AODV is conducted in order to investigate whether N-AODV effectively improves NTA with respect to AODV, and how this improvement affects both its effectiveness and efficiency. The experiment is executed within a simulation environment. The obtained results show that a MANET adopting N-AODV exploits higher NTA than a MANET adopting AODV. Moreover, the improved awareness impacts effectiveness and efficiency of routing activities because, in the long-run, it results in a lower need to activate the route discovery process for establishing new communications sessions.