Multi-scale Neural Style Transfer Based on Deep Semantic Matching

Abstract

Existing Neural Style Transfer (NST) algorithms do not migrate styles well to a reasonable location where the output image can render the correct spatial structure of the object being painted. We propose a deep semantic matching-based multi-scale (DSM-MS) neural style transfer method, which can achieve the reasonable transfer of styles guided by the prior spatial segmentation and illumination information of input images. First, according to real drawing process, before an artist decides how to paint a stroke, he/she needs to observe and then understand subjects, segmenting space into different regions, objects and structures and analyzing the illumination conditions on each object. To simulate the two visual cognition processes, we define a deep semantic space (DSS) and propose a method for calculating DSSs using manual image segmentation, automatic illumination estimation and convolutional neural network (CNN). Second, we define a loss function, named deep semantic loss, which uses DSS to guide reasonable style transfer. Third, we propose a multi-scale optimization strategy for improving the efficiency of our method. Finally, we achieve an interdisciplinary application of our method for the first time–painterly rendering 3D scenes by neural style transfer. The experimental results show that our method can synthesize images in better original structures, with more reasonable placement of each styles and visual aesthetic feeling.