(cvpr2020) reading of learning texture transformer network for image super resolution

Learning texture transformer network for image super-resolution

TTSR

This model can effectively search texture information from the reference image , Maximize the use of texture information of the reference image , And migrate to the super-resolution results .

Abstract ：SR It is to restore the real texture from the low resolution image . Some recent methods use high-resolution images as reference images , So that some related textures can be transferred to low resolution images . But some methods ignore the attention mechanism , Transfer high-resolution textures from the reference image . In this paper , The author puts forward TTSR The Internet , among LR And reference image respectively Transformer Queries and keys in .TTSR It contains four parts , Contains a learnable texture extractor 、 Correlation embedding module 、 Hard attention module for texture transmission 、 Soft attention module for texture synthesis . This design encourages cross LR Joint feature learning with reference image , The deep features can be found through the attention mechanism , Therefore, it can transfer texture features well . The proposed network can be Across scales Stack by .

introduction

Image hypersegmentation is usually divided into two categories , One is super-resolution reconstruction of single image （SISR）, The second is based on reference images