<![CDATA[Conventional pearl quality assessment remains heavily reliant on manual visual inspection, which is subjective and inconsistent. This study develops PearlVision AI, an automated system for grading Lombok pearls using morphological feature extraction and ensemble learning. The dataset comprises 361 South Sea pearl images (Pinctada maxima) labeled into three commercial grades: A (n=120), AA (n=120), and AAA (n=120). The proposed pipeline integrates hybrid segmentation (Hough Circle Transform + Convex Hull) for robust object isolation, extraction of four geometric descriptors (circularity, eccentricity, area, perimeter), and comparative evaluation of four classification algorithms: Random Forest, Gradient Boosting, K-Nearest Neighbor, and SVM (RBF). Results demonstrate that Random Forest achieved optimal performance with a test accuracy of 97.22% and a 5-fold cross-validation score of 91.68%, consistently maintaining precision, recall, and F1-score >0.95 across all grade classes. Feature importance analysis revealed that size-related features (area and perimeter) contributed more significantly to class discrimination than shape-based metrics (circularity), reflecting the natural correlation between pearl diameter and commercial value in this dataset. With an inference time of <0.5 seconds per image, PearlVision AI offers an objective, efficient, and reproducible solution for reducing manual grading bias and enhancing quality control consistency in the pearl industry]]>
Copyrights © 2025
