Recurrent Multi-scale Approximation-Guided Network for Single Image Super-Resolution

Single-image super-resolution (SISR) is an essential topic in computer vision applications. However, most CNN-based SISR approaches directly learn the relationship between low- and high-resolution images while ignoring the contextual texture and detail fidelity to explore super-resolution; thus, they hinder the representational power of CNNs and lead to the unrealistic, distorted reconstruction of edges and textures in the images. In this study, we propose a novel recurrent structure preservation mechanism with the integration and innovative use of multi-scale wavelet transform, Recurrent Multiscale Approximation-guided Network (RMANet) , to recursively process the low-frequency and high-frequency sub-networks at each level separately. Unlike traditional wavelet transform, we propose a novel Approximation Level Preservation (ALP) architecture to import and learn the low-frequency sub-networks at each level. Through proposed Approximation level fusion (ALF) and inverse wavelet transform, rich image structures of low frequency at each level can be recursively restored and greatly preserved with the combination of ALP at each level. In addition, a novel low-frequency to high-frequency detail enhancement (DE) mechanism is also proposed to solve the problem of detail distortion in high-frequency networks by transmitting low-frequency information to the high-frequency network. Finally, a joint loss function is used to balance low-frequency and high-frequency information with different degrees of fusion. In addition to correct restoration, image details are further enhanced by tuning different hyperparameters during training. Compared with the state-of-the-art approaches, the experimental results on synthetic and real datasets demonstrate that the proposed RMANet achieves better performance in visual presentation, especially in image edges and texture details.