Lightweight feature separation, fusion and optimization networks for accurate image super-resolution

Recently, single-image super-resolution (SISR) methods based on deep learning have demonstrated great superiority by deepening or widening the network. However, excessive network layers will not only weaken the information flow during training process, but also increase the storage load and computation cost in practical application. To achieve a better trade-off between model efficiency and accuracy, we propose a lightweight feature separation, fusion and optimization network (SFON) for SISR. For the architecture, we design an efficient feature separation, fusion and optimization block (SFOB) to effectively capture the local cross-level features through successive channel splitting and concatenation first, and then refine them with an improved channel attention mechanism. We also adopt a MAE pooling-based feature optimization and fusion block (MAE-FOFB) to enhance the distinction and utilization of global multi-level features extracted from every SFOB. For the loss function, except for L1 loss, the structural similarity (SSIM) loss is additionally introduced to fine-tune the network, which helps to bring a slight improvement in accuracy. Moreover, we develop a variant of SFON (SFON-P) by applying progressive reconstruction strategy to further boost performance. Extensive experiments show that both SFON and SFON-P achieve favorable reconstruction accuracy against other state-of-the-art lightweight models with relatively low model complexity.