PNRR PHYCOMOX

---

Bioretics won a PNRR Italian National Project denominated PHYsically COnstrained MOdeling with few eXamples (PHYCOMOX), CUP J13D24000090004

The project aims to extend, both theoretically and practically, an already operational system based on Machine & Deep Learning for non-destructive quality control on industrial sorting lines for fruit and vegetables (multi-class semantic segmentation on multi-spectral, multi-view, and multi-temporal images of fast-rotating and translating objects). The physical nature of the growth process leads to high intra-/inter-class variability in fruit, both temporally and geographically. Moreover, the lack of a shared domain definition prevents commercial specifications from being formally defined with sufficient scientific rigor. The objective is to manage such variability while reducing both human resources needed for annotation and overall computational complexity. The planned activities include: dataset creation via a data acquisition system, definition of ontologies, training under limited or absent labeled data conditions, and model monitoring with drift analysis. We propose the use of approaches such as Deep Learning, Few-Shot Learning, Transfer Learning, generative Data Augmentation (NeRF or 3DGS), and Active Learning, in synergy with established techniques like SVM. The economic impact of the project can be measured across the entire value chain. Improved accuracy can be translated into increased sales and reduced waste, as well as measurable savings in annotation time and computational resources.

Bioretics won a PNRR Italian National Project denominated PHYsically COnstrained MOdeling with few eXamples (PHYCOMOX), CUP J13D24000090004