Real-time weight estimation of Zucchini using laser-based 3D matrix and deep learning

Abstract

Accurate and efficient weight-based grading systems are essential for maintaining quality standards and improving productivity in modern agriculture. However, traditional methods are often manual, time-consuming, and prone to inaccuracies. This study presents a machine vision-based system for automated Zucchini weight estimation using a cost-effective laser-assisted 3D imaging method combined with deep learning. A red line laser, aligned perpendicular to a moving conveyor, captures video frames of Zucchinis. Height deviations from the laser profile are analyzed to generate a 3D matrix for each Zucchini. These data are resized, augmented, and used via transfer learning as inputs for pre-trained deep learning models, VGG16, ResNet101, and MobileNetV2. The study also compares this laser-based volumetric approach with traditional RGB image-based estimation. Among the models, MobileNetV2 using 3D matrix input achieved the highest accuracy, with an R² of 0.96 ± 0.03, root mean squared error of 19.01 ± 6.29  g, and mean absolute percentage error of 7.04 ± 2.63 %. The 3D matrix-based approach significantly outperformed the image-only method in accuracy and processing speed. Nevertheless, limitations such as sensitivity to ambient light, surface moisture, and the need for controlled imaging conditions must be considered for real-world deployment. This research introduces a scalable, low-cost solution for real-time weight grading of Zucchinis, with the potential for adaptation to other irregularly shaped horticultural crops. Furthermore, the system architecture can be extended to assess other practical applications of postharvest traits, such as ripeness, surface damage, or frostbite. Future work will focus on validating the system across diverse environmental conditions and expanding its functionality for broader use in precision agriculture.