Edge AI for Produce Recognition

The growing availability of computational power and advanced machine learning techniques has accelerated the shift of artificial intelligence (AI) from centralized cloud-based systems toward edge computing. This thesis explores the application of edge AI within the retail sector, focusing on automated checkout systems where reliable, real-time product recognition is essential. Cloud-dependent solutions in this context suffer from latency, privacy risks, and integration challenges, whereas edge-based approaches offer a more efficient and secure alternative. The work presented here details the design and implementation of an AI-driven product recognition system for Datalogic checkout scanners. Starting from an initial pipeline capable of detecting both produce and packaged goods, the system was optimized to reduce software footprint, operate within hardware constraints, and enable continuous adaptation to new store conditions and products. Key components include data acquisition, color correction, background subtraction, embedding, dimensionality reduction, and probabilistic classification, all integrated into a unified edge-compatible framework. Beyond the recognition pipeline, a companion infrastructure — consisting of a frontend, backend, and remote server — was developed to support model updates, feedback loops, and seamless integration with existing point-of-sale systems. The results demonstrate the feasibility and benefits of edge AI for retail checkout, highlighting design strategies that balance accuracy, speed, memory efficiency, and long-term maintainability. This work contributes to both the practical deployment of edge-based retail AI solutions and the broader discourse on designing robust, adaptive, and user-friendly AI systems for resource-constrained environments.