Multi-scale large receptive field feature distillation network for lightweight infrared image super-resolution

Infrared imaging plays a pivotal role in applications such as remote sensing, unmanned aerial vehicles, and security surveillance. However, current infrared detectors have low resolution, and improving resolution from the hardware perspective is expensive. To address issues such as low resolution and significant loss of high-frequency information in infrared images, this study introduces a lightweight neural network model. Compared to existing super-resolution models, the proposed model exhibits lower complexity and is convenient for edge deployment, enabling real-time processing which extends the application scope of infrared imaging. This approach employs a novel architecture specially designed for the characteristics of infrared data, which significantly enhances the image clarity with a minimal increase in parameters, thereby making it suitable for resource-constrained real-time applications. Specifically, we design a lightweight multi-scale feature extraction module that utilizes group convolutions with varying kernel sizes across groups to realize efficient feature extraction. In addition, we introduce a multi-scale feature fusion module that adaptively integrates features harvested from preceding layers. To enhance the network’s ability to learn high-frequency information, we added a Fourier transform branch to the network and incorporated a frequency loss function into the loss function. Extensive experimental results demonstrate that our proposed approach achieves superior performance with fewer parameters compared to existing lightweight super-resolution methods. Our code is available at https://github.com/clelevo/MSLRSR.