AI-Based Visual Measurement System for Quality Assessment in Agrifood Production

The increasing level of automation promoted by the industry 4.0 paradigm is driving the adoption of intelligent measurement systems that can replace subjective, manual inspections in industrial processes. In agrifood production, quality assessment tasks such as determining shape suitability and evaluating maturity are still largely performed by human operators, resulting in limited repeatability and scalability.This paper presents an AI-based visual measurement system for automated quality assessment. In this system, deep learning is employed as a virtual sensing instrument within a vision-based inspection chain. The proposed approach formulates quality evaluation as a multi-label visual measurement problem and simultaneously estimates shape suitability and maturity condition from RGB images. Transfer learning with pretrained convolutional neural networks is used to infer quality-related attributes. A sigmoid-based output formulation and binary cross-entropy loss allow for the independent estimation of multiple measurands.The reliability of the system is validated using a Monte Carlo repeated hold-out protocol. This protocol is designed to assess the robustness and repeatability of the system across multiple dataset realizations. A series of comparative experiments were conducted on a range of pretrained architectures. These experiments revealed the stability of contemporary deep models in the context of the designated measurement task. Moreover, gradient-based activation mapping is utilized as a qualitative validation tool to ascertain that the inference process relies on physically meaningful object regions rather than background artifacts.This system is a reliable, AI-assisted visual measurement solution that aligns with Industry 4.0 requirements and is compatible with future integration into automated, in-line sorting systems for smart agrifood production.