Painting Style Transfer for Head Portraits using Convolutional Neural Networks

摘要

Head portraits are popular in traditional painting. Automating portraitrnpainting is challenging as the human visual system is sensitivernto the slightest irregularities in human faces. Applyingrngeneric painting techniques often deforms facial structures. Onrnthe other hand portrait painting techniques are mainly designed forrnthe graphite style and/or are based on image analogies; an examplernpainting as well as its original unpainted version are required. Thisrnlimits their domain of applicability. We present a new techniquernfor transferring the painting from a head portrait onto another. Unlikernprevious work our technique only requires the example paintingrnand is not restricted to a specific style. We impose novel spatial constraintsrnby locally transferring the color distributions of the examplernpainting. This better captures the painting texture and maintains thernintegrity of facial structures. We generate a solution through ConvolutionalrnNeural Networks and we present an extension to video.rnHere motion is exploited in a way to reduce temporal inconsistenciesrnand the shower-door effect. Our approach transfers the painting style while maintaining the input photograph identity. In addition itrnsignificantly reduces facial deformations over state of the art.