Subspace Capsule Network for Blurred Image Super-Resolution
Zuhai Qin, Yi Tang, Wei Wang, Dongmei Xu, Yu Rong Pan · 2021
Super resolution of blurred image can effectively improve image quality and machine recognition accuracy. Existing methods use universal prior information in the data to deblur, and rarely consider image semantic prior information. In the field of deep learning, using more prior information can achieve better results. Inspired by the powerful object semantic information representation ability of the capsule network, in this paper, a super-resolution method based on subspace capsule neural network is proposed. The blurred image and the clear image have different semantic distributions in the subspace capsule feature space. The MNIST data set is de blurred, and the clear image is used as the input of the high resolution self encoder, and the blurred image is used as the input of the low resolution self encoder. A mapping network is used to map the blurred image subspace capsule feature distribution to the clear image subspace capsule feature distribution, then it is input into the decoder of subspace capsule network trained by the original clear image to get the desired super-resolution result. It is proved that the proposed method can effectively improve the image pixel quality.