Multi-Scale Feature Modulation for Efficient Image Super-Resolution

To address the challenge of deploying super-resolution technology on resource-constrained devices, we propose an efficient image super-resolution algorithm using multi-scale feature modulation. The algorithm incorporates a multi-scale feature modulator, processing input features across different scales to manage them globally and identify key representative features. These convolutions enhance the network's capacity to learn local information effectively. Additionally, they facilitate channel shuffling, optimizing the network's adaptability to varying feature representations and improving overall efficiency in image super-resolution. Extensive experimental results show that the proposed model has 192.72K parameters and 43.2G floating-point operations, marking an 85.91% reduction in parameter count and an 86.39% reduction in floating-point operations compared to EDSR. Moreover, with comparable performance, our approach requires significantly less computational resources.