Accurate almond variety identification is essential for quality assurance and automated agricultural processing, yet manual inspection remains inefficient. Deep learning models, although effective, often require high computational resources and provide limited interpretability. This study proposes a lightweight machine vision framework based on selective handcrafted representation design for almond variety classification. Almond kernels are first isolated using channel-guided segmentation based on the b ∗ component of the Commission Internationale de l’Éclairage L*a*b* (CIELAB) color space. Complementary visual characteristics are modeled using rotation-invariant uniform Local Binary Patterns (LBP) for texture description and statistical chromatic features extracted from multiple color spaces, where Hue–Saturation–Value (HSV) provides the most discriminative representation. Instead of exhaustively combining descriptors, texture and color features are selectively integrated to construct a compact and informative feature space. Experiments on four almond varieties using stratified nested 10-fold cross-validation show that Logistic Regression achieves 98.97% accuracy, approaching the performance of deep learning models such as MobileNet and EfficientNetB0. Quantitative analysis and visualization (t-SNE) further support the improved separability and stability of the proposed representation. The results demonstrate that carefully engineered low-dimensional representations can achieve high accuracy while maintaining computational efficiency and interpretability, making the framework suitable for practical automated food inspection and sorting applications. • A lightweight framework identifies almond varieties using selective LBP–HSV features. • CIELAB b ⁎ channel-guided segmentation enables robust kernel isolation. • Hybrid features enhance linear separability, validated by t-SNE visualization. • Automated TPE and Successive halving ensure efficient hyperparameter optimization. • Logistic Regression achieves 99.04% accuracy, outperforming several deep learning models.
Hoang et al. (Fri,) studied this question.