Re-parameterized feature attention distillation network for efficient thermal image super-resolution

Deep learning based infrared image super-resolution algorithms have achieved impressive reconstruction performance, but these models have complex structures and cannot be applied to handheld infrared thermal imagers with limited computing resources and memory size. To overcome these challenges, we propose a re-parameterized feature attention distillation network, named RepFADN, a simple yet efficient thermal image super-resolution network. Specifically, based on the characteristics of infrared images, the use of heavy parameter techniques and shallow residual connections has improved the feature extraction ability of standard convolution, allowing the network to achieve maximum benefits while maintaining portability. The reasonable use of multi branch convolution with 1 × 1 kernel effectively reduces the complexity of feature distillation structure and the introduction of attention branch further improves the performance. Extensive benchmark data experiments show that the proposed RepFADN outperforms CNN-based lightweight and efficient super-resolution networks in terms of performance evaluation indicators. Compared with the most advanced efficient super-resolution networks based on self-attention mechanism, e.g. SRFormer_light, the number of network parameters and FLOPs are 4 × smaller, the memory overhead is 18 × smaller, the inference speed is 40 × faster, and the network performance is similar or even better. Code will be available at https://github.com/shenjun1994/RepFADN.git.