Non-Destructive Mangosteen Volume Estimation via Multi-View Instance Segmentation and Hybrid Geometric Modeling.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Precision Agriculture

Background:

• Accurate, non-destructive fruit volume estimation is vital for precision agriculture, impacting quality grading and yield prediction.
• Traditional methods struggle with complex fruit geometries like mangosteen due to irregular calyx structures.
• Existing vision-based techniques face limitations with asymmetric protrusions, hindering accurate volume modeling.

Purpose of the Study:

• To develop a novel framework for accurate, non-destructive mangosteen volume estimation using 2D imaging.
• To address the challenges posed by complex fruit geometry and irregular calyx in volume modeling.
• To compare the performance of single-view versus multi-view geometric models for fruit volume prediction.

Main Methods:

• Employed a You Only Look Once (YOLO)-based instance segmentation model to separate the fruit body from the calyx.
• Utilized multi-view instance segmentation and hybrid geometrical feature modeling for quantitative volume assessment.
• Trained and evaluated eight regression models, including single-view and multi-view geometric regressors, polynomial hybrids, and ellipsoid approximations.

Main Results:

• Excluding the calyx significantly improved model performance (R2<0.40 without vs. >0.64 with).
• Multi-view models consistently outperformed single-view models, increasing average predictive accuracy from R2=0.6493 to R2=0.7290.
• The best hybrid linear regression model, combining area features (As1.5, Ab1.5) and ellipsoid volume (Vellipsoid), achieved R2=0.7290, MAPE=16.04%, and RMSE=31.9 cm³.

Conclusions:

• The proposed hybrid model offers a low-cost, interpretable, and flexible solution for real-time fruit volume estimation.
• This approach effectively handles complex fruit geometries, overcoming limitations of traditional methods.
• The model is suitable for integration into automated sorting and grading systems in post-harvest processing.