PROJECT TYPE:

• Scientific research project

PROJECT DURATION:

• 1.10.2025. – 30.9.2029.

PROJECT LEADER AND TEAM MEMBERS: 

• Assoc. Prof. Diego Sušanj, PhD (principal investigator), Juraj Dobrila University of Pula, Faculty of Engineering
• Assist. Prof. Karlo Griparić, PhD, Juraj Dobrila University of Pula, Faculty of Engineering
• Prof. Peter Peer, PhD, University of Ljubljana, Faculty of Computer and Information Science
• Assist. Prof. Žiga Emeršič, PhD, University of Ljubljana, Faculty of Computer and Information Science
• Assist. Prof. Blaž Meden, PhD, University of Ljubljana, Faculty of Computer and Information Science
• Prof. Kristijan Lenac, PhD, University of Rijeka, Faculty of Engineering
• Assist. Marina Banov, University of Rijeka, Faculty of Engineering

SHORT PROJECT DESCRIPTION:

The BioPatchNet project develops a distributed inference architecture for biometric person recognition based on parallel processing of image segments. Conventional biometric systems frequently suffer from high latency, limited throughput and reduced robustness when exposed to real-world conditions. BioPatchNet introduces a segmented processing pipeline in which each image is decomposed into smaller regions that are processed independently by feature-extraction nodes, producing compact representations suitable for identification and verification. This reduces data transfer requirements and accelerates inference.

After feature extraction, a classifier ensemble evaluates the generated vectors and produces decisions that are combined through a consensus mechanism. This design reduces sensitivity to individual classifier errors and stabilizes the overall output. The project includes the development of models trained on image segments, the definition of feature-vector quality metrics, the implementation of multiple fusion strategies and the integration of all components into a unified distributed inference pipeline.

The system will be evaluated on large-scale biometric datasets, with dedicated assessment of latency, robustness, security and scalability. The expected outcome is an architecture capable of significantly lower latency compared to monolithic systems, while maintaining or improving recognition accuracy. The solution is applicable to smart-cities, security systems, healthcare infrastructures and environments that require fast and reliable biometric processing.

PROJECT OBJECTIVES:

The main objective is to design and validate the scalable distributed BioPatchNet system for biometric person recognition. The system integrates segment-level feature extraction, classifier ensembles and adaptive fusion strategies.

Specific objectives include:

1. Segment-level feature extraction and evaluation of feature-vector quality. The goal is to develop models that preserve discriminative power even when operating on small image patches.

2. Design and evaluation of classifier ensembles, comparing deterministic and stochastic decision-fusion methods and selecting the configuration with the lowest error rate.

3. Implementation of a distributed inference pipeline capable of real-time operation across multiple nodes, including analysis of network delay and synchronisation effects.

4. Ablation studies on large-scale biometric datasets to identify critical system components and optimise the overall architecture.

The project will deliver a functional prototype, open-source code and scientific publications. Expected contributions include reduced latency, higher robustness and stability, and a reference architecture relevant for both academia and industry.

PROJECT HOLDER: Juraj Dobrila University of Pula