基于深度学习的图像与视频风格化研究与实现

发布时间：2018-02-01 16:41

本文关键词： 深度学习图像风格化卷积神经网络迁移学习视频风格化　出处：《中国科学院大学(中国科学院工程管理与信息技术学院)》2017年硕士论文　论文类型：学位论文

【摘要】：深度学习作为当今人工智能领域最热门的方向,在图像识别、语义分割、自然语言处理等领域取得了重大突破,显示出强大的学习和处理能力,甚至在部分领域超过人类的表现。因此,一些研究人员将深度学习应用于图像生成任务中,并在图像风格化方面取得了很好的效果。图像风格化即将一张图像在保持内容不变的同时换成另一种风格,这使得普通人也能拍出艺术风格的照片。但是当前最好的图像风格化算法在网络训练、自拍图像处理、视频风格化等方面还有很多问题需要解决,本文针对这些问题提出了对应的解决方案。(1)本文设计了一种图像风格转换网络U-StyleNet,并采用迁移学习的方法解决了网络参数初始化导致输出图像有亮块或黑块的问题,同时加快了网络的训练速度。(2)本文提出了四种方案解决自拍图像风格化人脸处理不好的问题,即设置较大的风格缩放参数、全局光滑处理、局部光滑处理、精细的风格比重。(3)本文提出了一种快速视频风格化的方案,详细介绍了快速视频风格转换网络FV-StyleNet的双通道多尺度输入结构。这种设计参考了现有的光流计算网络和图像风格转换网络,并将两者结合在一起,在风格化的同时考虑两帧图像的连续性,实现了快速而稳定的视频风格化效果。同时,双通道多尺度输入的结构具有非常高的并行性,可以充分发掘GPU的并行计算能力。FV-StyleNet同样采用迁移学习的训练方式,首先训练一个光流网络,然后在此基础上训练视频风格转换网络。本文提出的U-StyleNet成功地解决了风格转换网络参数初始化问题,同时提高了训练速度。四种优化方案提高了自拍图像风格化的质量。快速视频风格化方案FV-StyleNet极大地加速了视频的风格化,为视频风格化的应用推广创造了条件。
[Abstract]:As the most popular direction in the field of artificial intelligence, depth learning has made a great breakthrough in image recognition, semantic segmentation, natural language processing and so on, showing a strong learning and processing ability. Even in some areas it outperforms human performance. Therefore, some researchers apply in-depth learning to image generation tasks. And in the image stylization aspect has obtained the very good effect, the image stylization namely one image while keeping the content invariable at the same time changes another style. However, the best image stylization algorithm in the network training, selfie image processing, video stylization and other aspects need to be solved. This paper proposes a corresponding solution to these problems.) this paper designs an image style conversion network U-StyleNet. The migration learning method is used to solve the problem that there are bright blocks or black blocks in the output images caused by the initialization of network parameters. At the same time, it speeds up the training speed of the network.) in this paper, we propose four solutions to solve the problem of bad face processing in stylized selfie images, that is, setting larger style scaling parameters and global smooth processing. This paper presents a fast video stylization scheme. The two-channel multi-scale input structure of fast video style conversion network (FV-StyleNet) is introduced in detail. This design refers to the existing optical flow computing network and image style conversion network. Combining the two methods, the continuity of the two frames is taken into account while the two frames are stylized, and the fast and stable video styling effect is realized. At the same time, the dual-channel multi-scale input structure has a very high parallelism. We can fully explore the parallel computing power of GPU. FV-StyleNet also adopts the training method of migration learning. First, we can train an optical flow network. Then the video style conversion network is trained on this basis. U-StyleNet proposed in this paper successfully solves the problem of initialization of style conversion network parameters. At the same time, the training speed is improved. Four optimization schemes improve the quality of the stylization of the selfie image. The fast video styling scheme FV-StyleNet greatly speeds up the stylization of the video. It creates conditions for the application and promotion of video stylization.
【学位授予单位】：中国科学院大学(中国科学院工程管理与信息技术学院)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP18

【相似文献】