Dictionary Learning for Image Style Transfer

Abstract

Deep image representations learned by CNNs have been shown to successfully separate image content from style, but the generic feature representations learned by CNNs are far from interpretable. Drawing upon recent developments in neural network inspired extensions of Dictionary Learning (DL) to multi-layer models, I present a framework using dictionary learning and sparse code representations to separate image content and style in natural images. This framework provides an interpretable decomposition of images into high and low frequency representations that can be used to generate new images by combining the style and content of any two arbitrary images. These decompositions and visualizations of convolutional filters provide novel insights into the image representations learned by convolutional neural networks and demonstrate the potential of DL and sparse code representations for image analysis and transformation.