From Handcrafted Descriptors to Deep Embeddings: A Comparative Study for Regional Hazelnut Image Classification

Automated regional hazelnut image classification is relevant for agricultural traceability, origin aware sorting, and decision support in food processing systems. This study presents a unified comparative benchmark for three class hazelnut image classification using Van, Ordu, and Giresun categories. The experimental framework integrates classical handcrafted descriptors, deep transfer learning models, and reduced space hybrid classical and quantum compatible classifiers. In the classical branch, HOG and RGB color histogram features were combined and evaluated using KNN and SVM based classifiers. In the deep learning branch, ResNet50 and EfficientNetB0 were trained through a two stage transfer learning strategy, and Grad-CAM was employed to provide qualitative interpretability. In the hybrid branch, 1280 dimensional embeddings extracted from the GAP layer of the best EfficientNetB0 model were compressed into low dimensional spaces for reduced space SVC and QSVC experiments. Under the present image level split protocol, the strongest handcrafted baseline, PCA + SVM, achieved 98.90% accuracy and 0.9888 macro-F1, while ResNet50 reached 99.45% accuracy and 0.9945 macro-F1. EfficientNetB0 achieved 100% test accuracy, while the reduced space classical SVC trained on compressed EfficientNetB0 embeddings attained 99.63% accuracy and 0.9961 macro F1 on the full test set. Under the fair subset protocol, classical SVC achieved 98.33% accuracy, whereas the best QSVC configuration reached 90.83% accuracy. This result should be interpreted as a strong within dataset benchmark rather than definitive evidence of object level or deployment level generalization. Grad-CAM visualizations further suggested that the deep models focused primarily on semantically meaningful hazelnut regions rather than on background patterns. The findings indicate that representation quality is a central factor in this task, while the reported near ceiling performance should be interpreted in light of the image level partitioning strategy