Automated system for sorting defective blueberries and generating extraction coordinates using RGB machine vision

Automatic fruit sorting is a critical need in the agricultural export industry due to the limitations and costs of manual inspection. For blueberries, superficial external defects decrease their commercial value and hinder compliance with international quality standards. This paper proposes an Edge Computing-based RGB machine vision system for the non-destructive detection of defective blueberries. Furthermore, it automatically generates highly accurate extraction coordinates for an XY Cartesian robot equipped with a pneumatic suction system. The methodology comprises four sequential stages: controlled image acquisition within a dark chamber, hybrid segmentation and classification using a lightweight deep learning architecture (MobileNetV3), affine transformation for precise conversion of pixel coordinates to real-world millimeters, and the generation of synchronized kinematic movement instructions for the robotic system. Experimental validation was rigorously performed on an independent set of 100 blueberry samples. The implemented model achieved a robust classification accuracy of 95%. Additionally, the spatial transformation yielded a maximum localization error of less than 1.5 mm, perfectly aligning with the suction effector's tolerance. This was achieved while maintaining an efficient total processing latency of under 3 ms per fruit. These results demonstrate that the proposed mechatronic architecture constitutes an efficient, scalable, and low-cost solution for automated visual inspection and physical robotic handling of blueberries in continuous post-harvest production lines.