Abstarct: Pneumonia is a leading cause of infectious mortality worldwide, particularly affecting children and the elderly, necessitating rapid and accurate diagnosis. Chest radiography serves as the primary tool for clinical diagnosis. However, its manual interpretation is dependent on the radiologist's experience and is inherently prone to error. This research aims to present a novel and efficient hybrid deep learning model for the automated classification of three categories COVID-19, pneumonia, and normal from X-ray images, designed for deployment in resource-limited settings. In this study, a novel hybrid model is proposed, baxsed on the purposeful integration of the ResNet network, as an extractor of local hierarchical features, and the Swin Transformer V2 Tiny, as a modeler of global dependencies. The proposed model utilizes a frozen backbone of ResNet-50/101, pre-trained on ImageNet, which preserves the spatial resolution of feature maps by applying dilated convolutions in its deeper laxyers. This approach effectively extracts multi-scale hierarchical features from different levels of the image. To overcome the limitation of CNN models in modeling long-range dependencies, the extracted features, after channel reduction and the application of a Squeeze-and-Excitation(SE) block, are fed into two parallel Swin Transformer V2 Tiny models via a projection module. Only the last laxyer of the transformers is fine-tuned, resulting in training only approximately 25% of the parameters, thereby ensuring high computational efficiency. Finally, an intelligent fusion of intermediate and high-level features is performed through a bidirectional cross-attention mechanism and an MLP classifier to produce the final class prediction. This hybrid architecture combines the strengths of CNNs in extracting local and fine-grained features with the capability of transformers in modeling global and contextual relationships. Experiments conducted on the standard Cohen dataset demonstrated that the model with ResNet-50 achieved an accuracy of 94.49%, a mean F1-Score above 94%, and, most importantly, a mean AUC of 99.1%, which is on par with the current state-of-the-art. By maintaining high accuracy while significantly reducing computational complexity, the proposed model holds substantial potential for deployment in clinical settings with limited resources and represents an effective step towards the development of intelligent computer-aided diagnostic tools for pneumonia.