Robust Produce Recognition for Automated Retail: Data Augmentation and Domain Adaptation

In the context of automated product recognition systems, companies like Datalogic must ensure that their AI models remain accurate over time, even as acquisition conditions change. The main goal of this internship was to improve the robustness of the recognition pipeline and mitigate the performance drop due to domain shift. At the pre-classification stage, the product detection network was adapted to handle data acquired with a new scanner through a redesigned preprocessing phase. At the classification stage, a data augmentation strategy was added to the embedder's training pipeline to account for the fact that items in the new images appeared larger. As a result, the model maintained a Top-4 accuracy of 84\% when evaluated on items at different scales (+10 points compared to the baseline). Thanks to these improvements, Datalogic showcased a fully functional demo of its smart scanner at NRF 2025.