Abstract

The field of single-image super-resolution (SISR) has made considerable progress with the emergence of deep convolutional neural networks, where residual learning techniques have played a crucial role in enhancing reconstruction quality. Among these methods, SwinIR, a transformer-based network, demonstrates remarkable performance by leveraging hierarchical self-attention mechanisms to effectively capture both fine-grained local structures and broader global contextual dependencies. However, enhancing image quality while maintaining computational efficiency remains a key challenge.
To address the limitations in capturing diverse spatial features without increasing architectural overhead, we propose EMS network, an enhanced multiscale SISR framework that draws inspiration from the inception module to refine feature extraction across multiple scales. Our network design adopts the underlying principle of parallel multi-scale feature extraction from the inception module, where several convolutional layers with different receptive fields operate to capture spatial features at multiple scales. Our method maintains the lightweight design of the network while broadening its receptive field, allowing it to surpass existing state-of-the-art methods in both efficiency and reconstruction quality. Quantitative and qualitative evaluations demonstrate that enhanced multiscale network consistently outperforms state-of-the-art SISR models on standard benchmark datasets, delivering improved visual quality and superior quantitative performance with minimal impact on computational complexity.