Projects

Ongoing

Anomaly Detection

The project develops image-based anomaly detection methods for industrial quality control, with particular attention to scarce defective data, explainability, and distributed deployment. It combines synthetic defect generation, guided-attention architectures, and lightweight diffusion models to support anomaly classification, localization, and segmentation. Experimental results on MVTec AD show that generated defects improve recognition of real anomalies, while the proposed models provide competitive accuracy with reduced reliance on external datasets and large pretrained architectures.

Ongoing

Autonomous Driving

The project develops computer-vision and deep-learning methods for perception in autonomous and assisted-driving systems. Its activities focus on monocular depth estimation, pedestrian and cyclist distance assessment, semantic segmentation, and explainable scene understanding using low-cost RGB cameras. The proposed approaches improve depth accuracy and visual interpretability while reducing reliance on expensive sensors such as LiDAR.

Completed

Edge AI Technologies for Optimised Performance Embedded Processing

EdgeAI develops secure, scalable, energy-efficient, and trustworthy artificial-intelligence technologies for deployment across the edge continuum, from embedded devices to distributed edge infrastructures. The project integrates hardware/software co-design, model optimisation, explainability, and real-time processing in industrial demonstrators spanning digital industry, energy, agri-food, mobility, and digital society.

Completed

Granulometry

The project develops image-based methods for automatic granulometry, enabling contactless estimation of particle-size distributions for industrial and scientific applications. It combines neural processing, synthetic virtual environments, stereo vision, metric 3D reconstruction, and computational-intelligence techniques to estimate particle length, width, and thickness. The proposed systems reduce the limitations of manual sampling and conveyor-based inspection while achieving millimetric and sub-millimetric accuracy in realistic inline conditions.

Completed

Innovative Poplar Low Density Structural Panel

The I-PAN project developed an environmentally sustainable, lightweight engineered-wood panel composed of approximately 50% recycled wood and 50% poplar wood, including underused material from the upper part of the tree. The project combined innovations in strand production, drying, resin formulation, mat forming, and inline vision-based quality control. Image processing, stereo vision, and computational intelligence enabled automatic monitoring of strand size and quality, supporting lower density, reduced adhesive consumption, improved process control, and lower environmental impact.

Completed

Inspection of Particle Boards

The project develops automatic visual-inspection methods for detecting printing and mechanical defects on melamine-laminated particle boards. Its key idea is to recover the repetitive surface pattern directly from each board, reconstruct an ideal defect-free reference, and compare it with the acquired image. The proposed approaches achieved 96.6% overall defect-detection accuracy, while a genetic pattern-extraction method identified candidate patterns in up to 98% of the evaluated panels.

Completed

Laser Cut Inspection

The project develops image-based methods for automatic quality monitoring of metal laser cutting processes, in research carried out in collaboration with TRUMPF. A camera observes the jet of sparks produced beneath the cutting area, while image processing and neural classifiers relate its shape to defects such as excessive drag, discontinuities, and residual burrs. The resulting system achieved a mean frame-classification error of about 0.18%, demonstrating the feasibility of accurate near-real-time inspection without directly examining every finished edge.

Completed

Laser Weld Inspection

The project develops automated, real-time quality monitoring methods for laser welding of automotive components. The research was performed in collaboration with Philips and FIAT, combining industrial expertise with signal processing, feature selection, and compact neural classifiers. The system detects penetration problems, coupling misalignment, porosity, and laser-source faults while maintaining a computational cost compatible with industrial monitoring.

Ongoing

Machine Learning for DDoS Detection

The project develops adaptive machine-learning and deep-learning methods for detecting Distributed Denial of Service attacks in conventional, IoT, and Industrial IoT networks. Its research spans automatically configured neural architectures, transfer learning for evolving threats, and federated learning with resource- and data-aware client selection. Across multiple cybersecurity benchmarks, the proposed methods achieve high binary and multi-class detection accuracy while improving adaptability, privacy, and deployment efficiency.

Completed

Volume Estimation

The project develops a low-cost, contactless method for estimating object volume from a single synchronized two-view acquisition. A reduced 3D point cloud is reconstructed, summarized through geometric features, and processed by a neural network that corrects the initial convex-hull approximation. Experiments on 52 objects showed mean errors of 1.4% for parallelepiped-shaped objects and below 1% for cylindrical, spherical, and mixed-shaped objects.

Completed

Wood Classification

The project develops contactless and automatic methods for classifying wood species in industrial environments through fluorescence spectroscopy and computational intelligence. A laser excites the wood surface, while spectral features are extracted and classified using statistical and neural models, supporting up to 21 wood types. The proposed systems achieve classification errors as low as 6.4% for the full 21-species problem and operate with processing times compatible with real-time production.

Completed

Photovoltaic Energy Prediction

The project develops machine-learning methods for predicting photovoltaic energy production under changing weather, operating, and environmental conditions. It combines public meteorological data, computational-intelligence models, transfer learning, and digital twins to support both conventional and third-generation photovoltaic technologies. The proposed approaches improve prediction when only limited data are available and can also help identify performance degradation and schedule maintenance.

Ongoing

Wildfire Smoke Detection

The project develops low-cost computer-vision and computational-intelligence methods for early wildfire smoke detection from long-range video acquired by standard visible-light cameras. It analyses motion, colour, growth, upward movement, edge softness, and plume-shape irregularity to distinguish smoke from ordinary environmental changes. A complementary virtual environment generates realistic synthetic smoke under wind, fog, low illumination, and noise, improving training when real wildfire data are scarce.

Ongoing

1D Physiological Signals: ECG and PPG

The project investigates ECG- and PPG-based computational methods for biometric recognition, continuous authentication, and wearable health applications. It combines compact binary templates, correlation-based matching, deep neural networks, and cross-signal interoperability to recognize individuals from their cardiac activity with limited user cooperation. Results include fast HeartCode matching, deep ECG identification above 90% with short multi-lead signals, and MultiCardioNet equal-error rates of 2.15% for ECG, 2.42% for PPG, and 4.76% across the two sensing modalities.

Completed

3D Latent and Ancient Fingerprints

The project develops a low-cost, contactless two-view system for acquiring and reconstructing latent fingerprints impressed on fragile clay artworks. Stereo matching, triangulation, surface estimation, and texture mapping produce metric 3D fingerprint models that are less affected by perspective distortion and acquisition viewpoint. The method was validated on 275 image pairs from an artwork attributed to Antonio Canova, achieving a calibration error of 0.04 mm and a reconstruction error of approximately 0.5 mm.

Completed

ABC Gates for Europe

The ABC4EU – ABC Gates for Europe project developed and validated harmonised Automated Border Control solutions for faster, more secure, and less intrusive border crossings across Europe. The research combined second-generation e-passports, face and fingerprint recognition, multimodal biometric fusion, anti-spoofing, mobile border-control tools, and privacy-by-design principles. Pilot systems were tested at airport, seaport, and land-border sites, while dedicated biometric methods improved recognition accuracy under the non-ideal conditions encountered at operational e-Gates.

Completed

COntactlesS Multibiometric mObile System in the wild

COSMOS developed a unified framework for contactless multibiometric recognition in unconstrained environments, using smartphones, tablets and mobile cameras. The project combined complementary physiological and behavioural traits—such as face, iris, fingerprint, palmprint, ear, gait and gaze—through context-adaptive acquisition, matching and fusion. Its research addressed recognition accuracy, usability, computational efficiency, spoof resistance and privacy protection for mobile and real-world applications.

Ongoing

Iris Deidentification

Iris Deidentification addresses the privacy risks created by high-resolution facial images published on websites and social networks, from which iris patterns may be extracted for biometric recognition. The project developed a generative adversarial network capable of replacing identity-related iris texture while preserving the natural appearance and overall utility of the original portrait. It also introduced the I-SOCIAL-DB dataset and evaluation protocols for studying ocular recognition and measuring the effectiveness and visual realism of iris deidentification.

Ongoing

Privacy in Biometrics

Privacy in Biometrics investigates methods for using biometric recognition without exposing the permanent and sensitive characteristics of individuals. The research combines multimodal biometrics, secure sketches, cryptographic identifiers and homomorphic encryption to protect templates during storage, transmission and matching. The resulting approaches enable accurate authentication while limiting information disclosure, preventing cross-database tracking and supporting revocable, privacy-aware biometric systems.

Ongoing

Soft Biometrics

Soft Biometrics investigates the automatic extraction of personal characteristics—such as age and body weight—from facial images and video sequences acquired in non-cooperative, real-world conditions. The research combines image processing, pretrained deep neural networks, feature fusion and computational intelligence to obtain unobtrusive and computationally efficient estimates. These traits can support identification, surveillance, forensic analysis and human–machine interaction, either as preliminary filters or as complementary information in multimodal biometric systems.

Completed

Touchless 2D Fingerprint Recognition

Touchless 2D Fingerprint Recognition investigates biometric systems that capture fingertip images with conventional cameras, webcams, or smartphones without requiring contact with a sensor. The research develops dedicated methods for image-quality assessment, ridge-pattern enhancement, geometric correction, feature extraction, and matching under variations in pose, distance, focus, illumination, and motion. It also explores deep-learning techniques for extracting fine Level-3 details, such as sweat pores, enabling more hygienic, usable, and widely deployable fingerprint recognition.

Ongoing

Touchless 2D Palmprint Recognition

Touchless 2D Palmprint Recognition develops biometric methods that identify individuals from photographs of the palm acquired without physical contact, using cameras or mobile devices. The research addresses variations in hand position, scale, rotation, illumination and image quality through robust preprocessing, palm-specific descriptors and deep neural networks. It also combines palmprint and inner-finger texture extracted from the same hand image, improving recognition without requiring additional sensors or acquisitions.

Completed

Touchless 3D Fingerprint Recognition

Touchless 3D Fingerprint Recognition develops biometric systems that reconstruct the three-dimensional shape and ridge texture of a fingertip from images captured without physical contact. Multi-camera acquisition, geometric reconstruction, unwrapping, quality assessment and rotation compensation produce touch-compatible fingerprints that remain robust to finger position, perspective and environmental variations. The resulting technology improves hygiene, usability and interoperability while supporting accurate, on-the-move fingerprint verification in less-constrained environments.

Completed

Touchless 3D Palmprint Recognition

Touchless 3D Palmprint Recognition develops low-cost biometric systems that identify individuals from the three-dimensional shape and texture of the palm without requiring contact with a sensor. A two-camera acquisition setup and LED illumination reconstruct a metric 3D palm model, compensating for variations in hand pose, orientation, and distance. The approach combines 2D texture and 3D geometric features to achieve accurate, fast, hygienic, and user-friendly recognition in less-constrained conditions.

Ongoing

Unconstrained Iris Recognition

Unconstrained Iris Recognition investigates biometric recognition from ocular images captured in uncontrolled environments and with limited or no user cooperation. The research develops deep-learning methods for quality assessment, iris and reflection segmentation, super-resolution, feature extraction, and matching under blur, occlusion, gaze deviation, reflections, compression, and variable illumination. It enables iris recognition from consumer cameras, mobile devices, websites, and social media while also addressing computational efficiency, privacy risks, and biometric deidentification.

Ongoing

Biliary Atresia Detection

The project investigates a smartphone-assisted method for the preliminary detection of biliary atresia in newborns through stool-colour analysis. It combines colour clustering, limited user interaction, RGB/YUV feature extraction, and nearest-neighbour classification to distinguish normal from acholic stools while reducing sensitivity to shadows, reflections, and inaccurate pixel selection. Initial experiments on images acquired under uncontrolled conditions achieved 100% accuracy for the binary normal-versus-acholic task, although validation was limited to a small dataset.

Ongoing

Leukemia Detection

Acute Lymphoblastic (or Lymphocytic) Leukemia (ALL) is a disease that affects the blood cells, can spread rapidly throughout the body, and may result in fatal consequences if not detected at an early stage. One of the techniques routinely used to diagnose ALL consists in analyzing White Blood Cells (WBC) present in peripheral blood samples to look for malformations or abnormalities. Such malformations may be an indicator of lymphoblasts, which naturally occur in the bone marrow. However, an elevated number of WBCs with lymphoblast characteristics may be a sign of ALL.

Ongoing

X-Ray-based COVID-19 Assessment

The project develops a deep-learning pipeline for assessing COVID-19 and viral pneumonia from chest X-ray images using an embedded point-of-care system. It combines lung segmentation, reinforcement-learning-based self-augmentation, 3D feature generation, and non-local attention to classify normal, COVID-19, and viral-pneumonia cases. On the evaluated public radiography dataset, the method achieved approximately 98% accuracy, sensitivity, precision, F1 score, and specificity.