| Texture synthesis refers to the generation of large-scale textures based on given tex-ture exemplar,with the visual features and structure preserved.Compared with the sta-tionary texture,the main challenge of non-stationary texture synthesis is how to expand the global structure reasonably,such as the key contour lines and geometric shapes(the veins in leaf texture)and distribution trends of texture elements(the spots in feather tex-ture).In addition,considering the application of texture synthesis,such as the structure transfer,this paper focuses on the structural properties of non-stationary exemplars in texture synthesis.The existing non-stationary texture synthesis methods mainly suffer from the following two problems: 1)The traditional neighborhood-based matching meth-ods fail to obtain the structural attributes based on the small neighborhoods,while the(de)convolution and up/down sampling step in the convolution neural network are local operations,unable to capture the large-scale structure.2)Considering the needs of struc-tural transfer,the structural analogy and style transfer methods in image generation tasks can not handle the texture structure well,and thus can not achieve structurally controllable texture synthesis.Aiming at the two problems related to the above structural properties,this paper pro-poses two non-stationary texture synthesis methods from the perspective of reasonable expansion of the global structure and controllable synthesis respectively:· Hidden layer Fourier convolution for non-stationary texture synthesis.The main contributions of this method are as follows: 1)A texture synthesis method based on Fourier convolution operator is proposed,using the association between the global receptive field of frequency domain and the large-scale structure of the non-stationary exemplar,to im-prove the structural accuracy of the synthesized texture.2)Based on the dual relationship between the frequency domain and the spacial domain,a dual-domain branch is built to take into account the global structure and local texture details.3)Extensive comparative experiments and multi-level ablation studies demonstrate the effectiveness of our method on non-stationary texture exemplars.· A structure controllable method for non-stationary texture synthesis.The main con-tributions of this method are as follows: 1)A texture synthesis method including structure analysis and control is proposed,which aims to solve the large-scale structural challenges in non-stationary synthesis tasks.2)An explicit structure analysis approach is proposed to capture key structure-aware properties in textures,without relying on auxiliary net-works or semantic segmentation inputs.3)A novel structure-based synthesis model is constructed and a multi-scale perceptual loss is proposed to learn an accurate mapping from structural properties to texture exemplars. |
