Automated Variety Identification and Polishing Status Classification of Tur Dal Grains Using Computer Vision and Machine Learning

Abstract

Manual grading of tur dal (Cajanus cajan) remains subjective, inconsistent, and poorly scalable, underscoring the need for objective, image-based inspection. This study develops an integrated computer vision–machine learning framework for simultaneous variety identification and polishing-status classification of tur dal grains. Twenty-five morphometric, colorimetric, surface-optical, and defect-related features were extracted from 1,500 grain images (five varieties, two polishing states) and validated statistically (one-way ANOVA, chi-square, p < .001) before training a multi-layer perceptron classifier. On a held-out test set (n = 300), the model achieved 99.50% accuracy, 0.0006 loss, and precision/recall/F1 nearly equalling to 1.00 across grades, with ROC–AUC approaching 1.00; permutation importance identified weevil damage, colour uniformity, and broken-grain percentage as dominant predictors. Findings confirm feasibility under controlled imaging conditions, though generalisability to field-acquired, variably lit samples remains untested. The study's originality lies in unifying three inspection tasks within one interpretable, statistically grounded pipeline for an understudied pulse crop.