Capsule Networkを用いた顔画像の超解像手法

抄録

In recent years, super-resolution using deep learning has attracted attention. Super-resolution is a technology for converting low-quality images to high-quality images. Super-resolution can be applied to the technology to identify a criminal from the video of a security camera. It is difficult to identify the criminal from raw images, because security cameras are low image quality to record long-term images. In this study, we propose a method to super-resolution human face images using Capsule network. Capsule Network represents input values and output values as vectors, which makes it possible to learn features between a positional relationship and an orientation of faces. Therefore, it can be expected to generate a face image of higher quality than Convolutional Neural Network (CNN). We employ the CelebA data set,which is collected about 200,000 face images, as the training data. The low quality image is generated from the original CelebA image. Capsule network is trained using original high quality images as outputs and low quality images as inputs. Experimental results show that the super-resolution method using capsule network generates high quality face image than CNN.