Non-Destructive Volume Estimation of Oranges for Factory Quality Control Using Computer Vision and Ensemble Machine...

Area of Science:

• Agricultural Technology
• Computer Vision
• Machine Learning

Background:

• Accurate volume estimation is vital for industrial quality control, particularly in the food and agriculture sectors.
• Traditional methods for volume estimation can be time-consuming and destructive.
• Developing non-destructive, precise volume prediction techniques is essential for efficient quality assessment.

Purpose of the Study:

• To develop a comprehensive, non-destructive method for predicting orange volume using machine learning and computer vision.
• To create a reliable pipeline for estimating fruit dimensions and predicting volume.
• To enhance industrial quality control processes in agriculture.

Main Methods:

• Utilized top and side views of oranges with a calibrated marker to estimate four key dimensions.
• Engineered features beyond basic geometry, including surface-area-to-volume ratios and shape-based descriptors.
• Trained and fine-tuned a machine learning model, specifically a Stacking Regressor, on a dataset of 150 oranges.

Main Results:

• The Stacking Regressor model achieved a high R2 score of 0.971, outperforming single-model benchmarks like LightGBM.
• The method demonstrated robustness against fruit variability by relying on fundamental physical characteristics.
• The approach is adaptable for various produce types, indicating broad applicability.

Conclusions:

• The developed method offers a precise and non-destructive solution for orange volume prediction.
• This technique supports real-time density calculation for automated defect detection and quality grading in factory settings.
• The study highlights the potential for advanced computer vision and machine learning in agricultural quality control.