Research activity

1. Unsupervised techniques for artificial intelligence
2. Analysis of biomedical images
3. Industrial and environmental informatics
4. Intrusion detection systtems
5. Biometric systems

1. Unsupervised, generative, and explainable artificial intelligence

The research activities focused on the design and development of innovative artificial intelligence methodologies based on unsupervised, generative, and explainable learning paradigms, with the aim of reducing the dependence on extensively labeled datasets, generating informative data representations and synthetic samples, and increasing the interpretability of artificial intelligence models. In particular, unsupervised learning techniques have been investigated to process large amounts of data without requiring each sample to be associated with a corresponding ground truth. This approach reduces one of the main constraints that typically limits the amount of data available for training and enables the development of adaptive learning algorithms capable of exploiting the intrinsic structure of the data while maintaining high accuracy and limited operational constraints. The research activities also addressed generative artificial intelligence techniques for data analysis, representation learning, data augmentation, and the simulation of heterogeneous acquisition conditions. In this context, Generative Adversarial Networks, Autoencoders, Variational Autoencoders, and more recently diffusion-based models have been studied and developed to learn compact and informative representations, generate realistic synthetic samples, and model transformations in the data domain. Particular attention was also devoted to Explainable Artificial Intelligence, with the aim of increasing the transparency and interpretability of the developed models. Innovative methods have been investigated for identifying the features and image regions that contribute to the decisions of artificial intelligence systems, for learning highly discriminative and interpretable low-dimensional representations, and for analyzing and explaining the transformations performed by generative models. The proposed methodologies have been successfully applied to signal and image processing in several application scenarios. In particular, innovative approaches based on Deep Learning and Convolutional Neural Networks have been developed for unsupervised representation learning in biometric recognition and biomedical image analysis. Generative Adversarial Networks and Explainable Artificial Intelligence have been investigated for modeling and interpreting the aging process in face images, while Autoencoders and explainable learning techniques have been used to extract compact, highly discriminative, and interpretable representations from general-purpose images. More recently, generative and explainable artificial intelligence methodologies have also been applied to industrial anomaly detection and synthetic defect generation.

2. Analysis of biomedical images

The research activities regarded the design and development of innovative methodologies and software algorithms based on multidimensional signal processing, image processing, and artificial intelligence techniques for the realization of automated and adaptive biomedical systems. The proposed methods enabled the development of high-accuracy medical decision support systems able to analyze and classify medical images acquired using both conventional and low-cost imaging devices, with the aim of supporting clinicians and, where appropriate, enabling their use also by non-expert personnel. The research mainly focused on histopathological images, radiographic images, and images captured using mobile devices.

Histopathological images
The research activities focused on the design and development of original methods based on Deep Learning, Convolutional Neural Networks, transfer learning, metric learning, and Explainable Artificial Intelligence for the high-accuracy classification of white blood cells and the detection of Acute Lymphoblastic Leukemia. Particular attention was devoted to the development of medical decision support systems capable of providing interpretable predictions and of highlighting the image regions and morphological characteristics that contribute to the diagnostic decision. The research considered low-dimensionality and limited-size image databases of white blood cells, investigating learning strategies aimed at improving the robustness and generalization capability of Deep Learning models. The activities also addressed preprocessing and image decomposition techniques enabling Deep Learning models to process high-resolution and Whole Slide Images for tumor detection. Furthermore, innovative transfer-learning approaches based on heterogeneous databases of histopathological images were investigated, experimentally evaluating the extent to which knowledge acquired from different tissues, acquisition conditions, and imaging domains can be transferred to improve the accuracy and robustness of Convolutional Neural Networks for tumor detection.
Radiographic images
The research focused on the design and development of innovative methods based on Deep Learning, Convolutional Neural Networks, Reinforcement Learning, and semantic segmentation for the analysis and classification of chest X-ray images, with particular attention to the detection of COVID-19. The proposed approaches aimed at providing fast, accurate, and minimally intrusive computer-aided diagnostic procedures that could complement conventional clinical assessment.
Images captured using mobile devices
The research focused on the design and development of original algorithms based on image processing and pattern recognition for the analysis of fecal images captured using mobile devices, with the purpose of detecting the presence of biliary atresia through a scarcely intrusive procedure. Particular attention was devoted to the realization of low-cost decision support tools that could also be used by non-expert personnel for preliminary screening.

Deep Learning for Acute Lymphoblastic Leukemia (ALL) detection

3. Industrial and environmental informatics

The research activities regarded the design and development of innovative analysis methodologies, original hardware systems, and novel software algorithms for the monitoring, prediction, measurement, and adaptive classification of features extracted from multidimensional signals in industrial and environmental applications. The proposed solutions enabled high-accuracy results to be obtained using low-cost acquisition technologies capable of operating at long distances, reducing operational and environmental constraints and enabling remote deployment and easier system management. In particular, the research activities have been carried out in the areas of industrial and environmental informatics.

Industrial informatics
Innovative algorithms and methods based on multidimensional signal processing, three-dimensional modeling, image processing, and computational intelligence techniques have been studied and developed for the quantitative and qualitative analysis of materials, products, and industrial processes using signals and images acquired at a distance. Specific methods based on Deep Learning, Convolutional Neural Networks, Generative Adversarial Networks, diffusion models, and Explainable Artificial Intelligence have been investigated and developed for industrial visual inspection and anomaly detection. The research addressed the analysis, generation, localization, and detection of defects in industrial products, with particular attention to synthetic defect generation, explainable anomaly detection, and the realization of models capable of highlighting the image regions that contribute to the identification of anomalous samples. The activities also considered distributed learning approaches for anomaly detection, in which the generation of synthetic defects and the learning process can be performed across multiple acquisition sites while reducing the need to directly share the original data. Moreover, innovative methodologies were developed for monitoring the production process of strand boards with low environmental impact and for estimating the volume of objects acquired using multiple-view imaging systems. The proposed approaches were also experimentally validated through the deployment of a prototype in an industrial production environment. The research activities also focused on the study and development of innovative pattern recognition approaches based on signal and image processing and artificial intelligence techniques for automotive and autonomous-driving applications. In this context, original methods based on Deep Learning and Convolutional Neural Networks were developed for monocular depth estimation, enabling the distance of objects in a scene to be estimated without requiring additional range sensors. The research also investigated image decomposition and patch-based representation strategies aimed at improving the effectiveness and robustness of monocular depth estimation methods. Furthermore, methods based on Convolutional Neural Networks were studied for the analysis of the driver's attention level, considering semantic segmentation techniques applied to images acquired by vehicle-mounted cameras as well as photoplethysmography-based approaches. Lastly, the research activities addressed the development of innovative Deep Learning-based methodologies for monitoring and estimating the health status of electrical and electronic components in new-generation electric vehicles.

Environmental informatics
Innovative algorithms and methods based on multidimensional signal processing, computational intelligence, and simulation techniques have been studied and developed for environmental monitoring and prediction systems. In particular, the research focused on advanced wildfire detection methods based on the analysis of frame sequences acquired at long distances using low-cost cameras, and on renewable-energy prediction using weather forecasts or information related to energy imbalance in smart grids. Privacy issues and privacy-preserving solutions for environmental monitoring systems were also investigated.

Wildfire smoke detection using computational intelligence

4. Intrusion detection systems

Artificial intelligence-based methods have been designed and developed for the adaptive detection and classification of cyberattacks from heterogeneous datasets containing network activity logs. The proposed techniques achieved high accuracy in distinguishing benign network traffic from malicious traffic, while also enabling the identification of different classes of intrusions. In particular, adaptive methods for sizing neural architectures have been designed and implemented to match the complexity of each specific dataset. These approaches have also been investigated in combination with knowledge transfer techniques, with the aim of improving the adaptability and generalization capability of intrusion detection systems across heterogeneous datasets and network environments. Moreover, innovative methods based on federated learning have been studied and developed for distributed intrusion detection, with particular attention to the adaptive selection of clients and training data. These approaches aim to improve the accuracy, robustness, and efficiency of attack detection while limiting the need to centrally collect and share network data.

5. Biometric systems

In the field of biometric systems, the research activities have regarded the design and realization of innovative solutions consisting in original hardware systems, software algorithms, and biometric methodologies for recognizing individuals in security applications using physiological traits and soft biometric features. The proposed methods enabled to reduce operational constraints of traditional biometric systems and allow a high-accuracy recognition using less-constrained and high-usability acquisition procedures. In particular, the research focused on touchless and less-constrained biometric systems and on high-usability touch-based systems.

Touchless and less-constrained biometric systems
In this context, the activities have been focused on innovative touchless and less-constrained acquisition procedures with a high usability, on new methods for the encoding of signal and images with a high level of noise, and on original methods for the comparisons of identities. In particular, different biometric traits have been studied.

Touchless and less-constrained fingerprint recognition
Innovative hardware systems, algorithms and biometric methods have been studied and realized for the recognition of fingerprints captured using less-constrained acquisition procedures, without the contact of the finger with the sensor. A particular focus has been given on the techniques for the processing of three-dimensional models for real-time identity comparisons and forensic applications. Methods for the extraction and analysis of features from samples captured using a single camera, mobile devices, as well as captured from heterogeneous sources, have been studied and realized. Methods for the quality estimation and the generation of synthetic of three-dimensional models have also been proposed.

Touchless and less-constrained palmprint recognition
Innovative hardware systems, algorithms and biometric methods have been studied and realized for the recognition of palmprints captured using touchless and less-constrained acquisition procedures, without the contact of the hand with any surface. A specific focus has been given to novel Deep Learning-based techniques using original Convolutional Neural Networks, designed to extract highly-discriminative biometric features from palmprint images, trained using innovative unsupervised procedures. Original techniques based on Convolutional Neural Networks for the fusion of palmprint and finger texture features have also been proposed to increase the accuracy of the biometric recognition. The research also focused on new techniques for processing three-dimensional models and using pattern recognition methods for comparing samples captured with different positions and orientations.

Touchless and less-constrained recognition using soft biometric traits
Innovative hardware systems and algorithms, based on multidimensional signal processing and computational intelligence techniques, have been studied and realized for the estimation of the weight of walking individuals in surveillance applications using touchless and un-obtrusive acquisition procedures. Multidimensional signal processing and computational intelligence techniques have also been used to estimate the age of the individuals using face images captured using un-obtrusive procedures. Explainable Artificial Intelligence techniques have been studied to analyze the learning process of Deep Learning-based methods that use Generative Adversarial Networks to synthetically age face images.

Less-constrained iris recognition
The research focused on studying the problems related to iris recognition and user tracking in security contexts, using public-domain images present on the internet and social media. To this purpose, the research activities also included the collection and publication of the first database, available to the public, of iris samples extracted from face images captured in less-constrained conditions and from images freely downloadable from the internet. The research acitivities included also the realization of an innovative methodology based on Generative Adversarial Networks to ensure the biometric anonymization of the iris in such images, at the same time maintaining a high visual realism of the obtained images.
- Iris deidentification with high visual realism
- I-SOCIAL-DB: Iris images collected from websites and social media

Highly-usable touch-based biometric systems
Recent advances in biometric recognition for Automated Border Control (ABC) systems have been studied along with emerging technologies. Innovative algorithms and original biometric methods based on multidimensional signal processing and computational intelligence techniques have been studied and realized for a more accurate, fast, and highly-usable touch-based fingerprint recognition in new-generation Automated Border Control systems. In these applications, the research has focused on multimodal biometric systems, on techniques for the classification of acquisition problems affecting fingerprint images, and on methods for biometric score normalization.

Advanced Border Control for Europe (ABC4EU)