Image Analogy with Gaussian Process

Abstract

Image analogy is the process of creating an image filter that precisely reflects the characteristics contained in the training data. Recently, the image analogy problem was generally handled by deep neural network (DNN) with the development of a deep learning technology. Generally, DNN suffers from a fatal problem in that it requires large amounts of data for training. However, as pairs of images with the same relationship are needed for an image analogy, it is hard to collect sufficient data for image analogy using DNN. In order to solve this problem, we propose an image analogy method using a Gaussian process. In this method, a Gaussian process regression is used instead of DNN regression to adjust the feature vectors which will be used in creating filtered image. Additionally, in order to accelerate the training speed of Gaussian process, we also propose novel sampling methods that select salient instances from a given dataset. Our experiment result demonstrates that the proposed image analogy method using a Gaussian process with salient instance sampling performs significantly better than DNN in environments with small dataset size.