Multi-scale Super-resolution Image Reconstruction based on Visual Feature and Scale Feature Transformation
Lili Liang, Di Miao, Meng Han · 2023
Aiming at the shortcomings of fixed scale and neglecting human visual characteristics of the existing deep learning-based super-resolution image reconstruction (SRIR) methods, this paper proposes a multi-scale SRIR model based on the visual feature and the scale feature transformation (SFT). Different from the traditional methods, the proposed model takes the scale factor as input parameter, and obtains the corresponding deep features by applying scale extension and SFT to the backbone network. Meanwhile a weight prediction module is constructed to dynamically learn the multi-scale up-sampling filters. By multiplying them with the deep features, the images with different super-resolutions are accordingly reconstructed. To further improve the visual effect of the reconstructed images, a frequency division-based loss function is constructed which matches the human visual perception well. Experimental results on several datasets such as Set14, Urban100 and General100 indicate that the proposed model can not only obtain super-resolution images at different scales, but also make the reconstructed image have more clear details and better visual effect.