Stylized Sketch Generation using Convolutional Networks

Abstract

The task of synthesizing sketches from photographs has been pursued with image processing methods and supervised learning based approaches. The former lack flexibility and the latter require large quantities of ground-truth data which is hard to obtain because of the manual effort required. We present a convolutional neural network based framework for sketch generation that does not require ground-truth data for training and produces various styles of sketches. The method combines simple analytic loss functions that correspond to characteristics of the sketch. The network is trained on and evaluated for human face images. Several stylized variations of sketches are obtained by varying the parameters of the loss functions. The paper also discusses the implicit abstraction afforded by the deep convolutional network approach which results in high quality sketch output.