Structure-preserving dynamic texture generation algorithm

摘要

A dynamic texture refers to a sequence of images with regular and periodic changes in appearance from one image to another. Typical dynamic textures include burning flames, waving flags and flowing water. Given a reference dynamic texture, dynamic texture generation is meant to generate new dynamic textures that are of similar appearance to the reference one but are more diversified in terms of dynamics (i.e., periodic pattern changes). Existing dynamic texture generation algorithms may suffer from partially losing appearance information of the reference dynamic texture, resulting in such problems as structural distortion and blurred edges in the generated dynamic textures. To address this issue, we propose in this paper a structure-preserving dynamic texture generation algorithm based on a new two-stream convolutional networks model with feature reuse strategy (TSCN-FR). The proposed TSCN-FR makes use of spatial co-occurrence feature matrices to capture the spatial structure correlation of dynamic textures, thus alleviating the structural distortion and disordered color patch distribution in the generation of dynamic textures. In addition, an edge-preserving regular term, which uses the edge features of the reference dynamic texture extracted by the LoG operator, is introduced in the training of TSCN-FR to better maintain the edge structure of the reference dynamic texture. The application of our proposed method to dynamic style transferring is also presented. Experimental results have shown that our method for dynamic texture generation improves the existing methods by a clear margin on the ability to keep appearance information of the reference dynamic texture.