Dynamic Image Super-resolution Network Based on Semantic Prior
Mingwei Yang, Xinying Xu, Lan Cheng, Zhe Zhang · 2021
In recent years, due to the vigorous development of deep learning, the super-resolution of a single image has made great progress, but there are still challenges in texture restoration. To recover natural textures, we build a dynamic convolutional network to model the diverse mapping relations. Faced with different low-resolution images, it can reconstruct high-resolution images through different mapping relationships. The parameters of dynamic layers are adaptively generated according to the semantic prior described by deep features of the input. In the convolution neural network, the deep features can often extract the regions related to the semantics of the image. according to different regions, our model can change the transmission parameters in the super-resolution network, to recover different texture details. Finally, the experimental results show that, compared with the most advanced SFT -GAN method, this method has better performance in texture restoration, especially for images of complex scenes.