Single Image Super-Resolution Based on Capsule Network
Dongmei Xu, Yi Tang, Zuhai Qin, Yu Rong Pan · 2021
Super-resolution reconstruction of a single image refers to the process of reconstructing a low-resolution image into a high-resolution image. Although image super-resolution has made good progress, the existing methods still rarely consider the semantic prior information of the image. In the field of deep learning, the more a prior information is used, the better the effect will be achieved. Based on the powerful object semantic information representation ability of capsule network, we propose a capsule neural network based on semantic information super-resolution. The mapping network is used to map the feature distribution of low-resolution image to the feature distribution of high-resolution image, and then input the decoder trained by high-resolution image to obtain the corresponding high-resolution image. The super-resolution task on data sets such as MNIST and set5 is visual display, which proves that our proposed method is effective.