基于改进AdaIN的图像风格迁移

Image Style Transfer Based on Improved AdaIN

图像风格迁移技术是指给定内容图和风格图,利用机器学习算法将内容图渲染成具有艺术风格的画作。针对传统图像风格迁移算法无法兼顾速度与生成图像质量等问题,基于AdaIN算法,提出AdaIN改进算法。在原始AdaIN网络中加入内容图像的深度信息计算模块,提取内容图像的深度图,通过将迁移图像数据与深度图归一化处理后的数据按元素相乘的方法,突出内容图的深度信息,使得输出的风格迁移图像各深度下具有不同的风格化程度。实验表明,相较于Gatys等多种传统风格迁移算法,AdaIN改进算法在运行时间上可降低约11%;不需要针对每种风格单独训练网络,避免了模型重复训练;内容图像深度信息得以保存,提高了图像渲染质量。

The technology of image style transfer is to render the content image into a painting with artistic style by using machine learning algorithm given the content image and style image.Aiming at the problem that the traditional image style migration algorithm can’t take into account the speed and the quality of the generated image,this paper proposes an improved AdaIN algorithm based on the AdaIN algorithm designed by Xun Huang et al.By adding the depth information calculation module of the content image into the original AdaIN network,the depth map of the content image is extracted.After the normalization of the migrated image data and the depth map,the data is processed by elements.The method of multiplication highlights the depth information of the content map,which makes the output style migration image with different stylization degree in different depth.Experimental results show that the improved algorithm described in this paper can reduce the running time by about 11%compared with many traditional style migration algorithms such as Gatys.It does not need to train the network separately for each style,avoiding the repeated training of the model.The depth information of the content image can be saved and the rendering quality of the image is therefore improved.