Enhanced Multi-Scale Network for Single Image Super-Resolution

The field of single-image super-resolution (SISR) has seen significant advancements with the emergence of deep convolutional neural networks, where residual learning techniques have contributed to notable improvements in reconstruction quality. Among these approaches, SwinIR [1], a Transformer-based model, has demonstrated impressive performance by leveraging hierarchical self-attention mechanisms to capture both local fine-grained structures and global contextual dependencies. However, improving image quality while maintaining computational efficiency remains a key challenge. To address this, we propose a multi-scale SwinIR inception-based network, an enhanced SISR framework that draws inspiration from the Inception module to refine feature extraction across multiple scales without introducing significant computational overhead due to the complexity of the network architecture. Instead of directly implementing the Inception module, we adopt its core idea of parallel multi-scale processing, where multiple convolutional layers with different receptive fields operate simultaneously to extract spatial features at varying scales. This strategic enhancement significantly improves PSNR over the original SwinIR model while increasing the parameter count by only 65K. Our model integrates hierarchical self-attention with multiscale feature extraction to strengthen the representation of structural details in low-resolution images. Experimental results demonstrate that EMS(Enhanced multi-scale network) consistently outperforms state-of-the-art SISR models across multiple benchmark datasets, delivering improved visual fidelity and superior quantitative performance without a significant increase in computational complexity.