Abstract
<title>Abstract</title> <p>Convolutional neural networks and vision transformers offer complementary strengths for image classification, but either paradigm alone can struggle to balance local feature discrimination and global context modeling on mid-scale datasets. This paper presents Informative Feature Fusion Hybrid Vision Transformer (IFF-HVT), a hybrid architecture that combines a ResNet-50 backbone, transformer-based context modeling, and an informative feature fusion module with cross-modal attention, sigmoid-gated fusion, and Squeeze-and-Excitation channel recalibration. The method projects convolutional features into token embeddings, models long-range dependencies with either global self-attention or shifted-window attention, and fuses convolutional and transformer representations before classification. A controlled comparison of eight candidate backbones identifies ResNet-50 as the best trade-off between accuracy and efficiency for the proposed framework. Experiments on CIFAR-10, CIFAR-100, MNIST, and SVHN show consistent gains over a ResNet-50 baseline, with the largest improvement on CIFAR-100 and test-loss reductions of 14 to 28 percent. Ablation studies show that cross-modal attention, gating, and channel recalibration each contribute to performance, and sensitivity analysis supports the robustness of the selected configuration. These results indicate that explicit fusion of local and global representations can improve accuracy while keeping computational cost close to that of a strong convolutional baseline.</p>