Produce Recognition: Deep Learning Methods for Robust and Efficient Deployment

Automatic recognition of fresh produce is a challenging computer vision task due to high visual variability, open-set operating conditions, and strict industrial constraints on latency and computational resources. Practical systems must operate reliably on embedded hardware while remaining robust to unseen categories, visual edge cases, and quantization effects. This thesis presents the development and analysis of an embedding-based produce recognition system carried out during an industrial internship at Datalogic. Building on an existing recognition pipeline, the work investigates the interplay between representation learning, quantization strategies, and open-set recognition in the context of real-world deployment. The thesis examines post-training quantization and Quantization-Aware Training for embedding-based models and introduces a training paradigm designed to preserve embedding consistency after quantization. In addition, hierarchical supervision strategies leveraging both coarse and fine-grained labels are explored to enhance representation learning. Finally, metric learning approaches based on angular margin losses are considered as an alternative framework for structuring the embedding space. Overall, the thesis focuses on aligning embedding learning objectives with the requirements of the final inference pipeline, with the goal of improving robustness and deployability under practical industrial constraints.