Artistic Image Style Transfer Based on the Octave Convolution Modulation

Currently, most image style transfer methods have achieved good stylized results. However, some methods lose a large number of style details during the style transfer process, and the overall style is significantly diluted, thereby leading to the phenomenon of style degradation. To alleviate this problem, we propose an artistic image style transfer based on the octave convolution modulation (OctAST) in this paper. Our OctAST decomposes an image into high-frequency and low-frequency features, which enables a better extraction of the style from the style image. Then, this method stylizes the high-frequency features and the low-frequency features respectively. We introduce the octave modulation module (OMM) during the process of image style transfer. The octave modulation module injects more flexible and complex style signals into content images, thereby enhancing the stylization effect of the images. In order to retain the original style color information, we propose the octave residual module (ORM). This paper demonstrates the effectiveness of the OctAST method through a large number of experiments.