Edge and Texture Enhanced Reference based Super-Resolution Network for Remote Sensing Images
摘要
The lack of high-frequency content and loss of important information in the low-resolution (LR) images has hindered the advancement of single-image super-resolution reconstruction (SRR). These limitations diverted the focus of researchers on an alternative, still effective method of reference based super-resolution reconstruction (RefSR). These models have shown promising results in reconstructing high-resolution (HR) images compared with state-of-the-art (SOTA) single-image SRR. However, most of the existing RefSR models rely on constraints such as the reference (Ref) image should have high similarity and being well aligned with that of a LR image. Moreover, they mainly focus on efficient texture transfer while leaving the edge reconstruction part unnoticed. To address this problem, we propose a RefSR model for remote sensing (RS) images, consisting of a dedicated texture enhancement (TE) module and a highly efficient edge enhancement (EE) module in the network architecture. These curated modules help to focus on the two most essential aspects of the SRR problem by extracting highly efficient edge and texture features. To eradicate the limitations of using a classification model like VGG-19 Net, we design an end-to-end trainable autoencoder network for texture feature extraction and transfer. To suppress the irrelevant details and promote the transfer of relevant features from the (Ref) and LR images, a deep feature attention module (DFAM) is proposed. The ablation study manifests that the DFAM enriches the network’s feature representation capability and enhances the model’s performance. The experimental results show that our proposed model significantly outperforms SOTA, achieving an improvement of approximately