ResNet-ViT-SVM a New Hybrid Architecture Proposal and Experimental Comparisons on Date Fruit

Abstract

Accurate classification of date fruit varieties is essential for quality control, intelligent sorting, and agricultural sustainability. This study proposes a novel hybrid deep learning framework, named ResNet-ViT-SVM, to classify nine different date fruit varieties with high precision. The dataset consists of 1658 images distributed across nine classes, captured under controlled and real-world conditions. The approach consists of three stages: (i) initial classification using a fine-tuned ResNet50 convolutional neural network, (ii) reclassification using a Vision Transformer (ViT), and (iii) fusion of deep features from both models, followed by final classification via Support Vector Machines (SVM). The novelty of our approach lies in this unique integration of CNN, Transformer, and SVM components in a three-stage pipeline. Experimental results show that the ResNet50 and ViT models individually achieved classification accuracies of 93.05 % and 95.47 %, respectively, while the proposed ResNet-ViT-SVM hybrid model significantly outperformed them, achieving up to 99.40 %. As part of the ablation study, the hybrid model achieved 100 % accuracy on laboratory images and 83.11 % accuracy on field images captured under natural orchard conditions, confirming its effectiveness across both controlled and real-world scenarios. These findings demonstrate that the hybrid architecture offers strong generalization capability across different data domains and represents a highly accurate, contactless, and automated solution for agricultural product classification tasks. © 2025 Elsevier B.V., All rights reserved.