Frequency-Separated Attention Network for Image Super-Resolution

The use of deep convolutional neural networks has significantly improved the performance of super-resolution. Employing deeper networks to enhance the non-linear mapping capability from low-resolution (LR) to high-resolution (HR) images has inadvertently weakened the information flow and disrupted long-term memory. Moreover, overly deep networks are challenging to train, thus failing to exhibit the expressive capability commensurate with their depth. High-frequency and low-frequency features in images play different roles in image super-resolution. Networks based on CNNs, which should focus more on high-frequency features, treat these two types of features equally. This results in redundant computations when processing low-frequency features and causes complex and detailed parts of the reconstructed images to appear as smooth as the background. To maintain long-term memory and focus more on the restoration of image details in networks with strong representational capabilities, we propose the Frequency-Separated Attention Network (FSANet), where dense connections ensure the full utilization of multi-level features. In the Feature Extraction Module (FEM), the use of the Res ASPP Module expands the network's receptive field without increasing its depth. To differentiate between high-frequency and low-frequency features within the network, we introduce the Feature-Separated Attention Block (FSAB). Furthermore, to enhance the quality of the restored images using heuristic features, we incorporate attention mechanisms into the Low-Frequency Attention Block (LFAB) and the High-Frequency Attention Block (HFAB) for processing low-frequency and high-frequency features, respectively. The proposed network outperforms the current state-of-the-art methods in tests on benchmark datasets.