基于深度学习的图像水墨风格渲染应用

发布时间：2018-08-03 07:25

【摘要】：随着信息技术,生物技术的快速发展,以及计算机运算性能的大幅提高,近两年来以深度学习(DL,Deep Learning)为代表的人工智能技术发展迅猛。深度学习相对于以前的机器学习,主要体现在神经网络的层数更深,神经元之间的连接更加广泛,能够自主学习到更多的隐藏特征,因此被广泛应用于文字,图像,语音等领域。在绘画方面,人类已经掌握了特有的绘画技巧,这可以被称为内容与风格的巧妙融合。计算机由于不具备人类的视觉神经系统,因此长期以来被认为不适用于艺术的创作。然而目前有德国学者提出利用深层的卷积神经网络模型来进行油画的创作,通过卷积神经网络提取油画的纹理特征,将此纹理融合在任意的内容图片上,生成了具有该油画风格的内容图片。本文在此基础上尝试将卷积神经网络的艺术风格算法应用在中国水墨画的创作上,设计和实现了 WEB服务器端程序,可以对任意图片进行水墨风格的渲染。WEB软件架构上采用B/S模式实现了逻辑层和表示层的分离,用户可以通过浏览器运行水墨风格渲染序,执行选择风格图片,内容图片,调整运行参数等操作。本文做的主要工作如下:1.分析了目前流行的深度学习框架和卷积神经网络模型,在评估了各深度学习框架优缺点和卷积神经网络结构模型的基础上,设计改进水墨风格渲染的算法。2.用Python编写了水墨风格渲染程序。将Caffe架构下的VGG,Illustration2Vec模型移植到Tensorflow架构,并且基于深度学习框架Tensorflow和卷积神经网络模型VGG-19,Illustration2Vec进行图像的水墨风格渲染,比较分析了不同模型以及同一模型不同网络层组合产生的图片效果,在分析过程中用Tensorboard可视化损失函数的梯度变化过程。3.基于Node,js搭建了图像水墨画风格渲染的服务器端程序,采用Express作为Web端框架,前端的页面效果通过HTML+CSS+JavaScript完成。
[Abstract]:With the rapid development of information technology, the rapid development of biotechnology and the dramatic improvement of computer computing performance, artificial intelligence, represented by DL (Deep Learning), has developed rapidly in the last two years. It has been widely used in the fields of words, images, phonetics, and so on. In painting, human beings have mastered the unique painting skills. This can be called the skillful fusion of content and style. Because the computer does not have the human visual nervous system, it has long been considered to be unsuitable. At present, the German scholars have proposed to use the deep convolution neural network model to create the oil painting, and extract the texture features of oil painting through the convolution neural network. The texture is fused on any content picture, and the content pictures with the oil painting style are generated. The WEB server side program is designed and implemented by the network art style algorithm in the creation of Chinese ink painting. The.WEB software architecture of any picture can be rendered by the B/S mode to separate the logic layer and the presentation layer. The user can run the ink style rendering sequence through the browser and execute the selection. The main work of this article is as follows: 1. the current popular deep learning framework and convolution neural network model are analyzed. Based on the evaluation of the advantages and disadvantages of the depth learning framework and the convolution neural network structure model, the design of the algorithm.2. for improving the style of the ink style is designed by Python. The ink style rendering program is written. The VGG, Illustration2Vec model under the Caffe architecture is transplanted into the Tensorflow architecture, and the ink style of the image is rendered based on the depth learning framework Tensorflow and the convolution neural network model VGG-19, and Illustration2Vec is used to analyze the different models and the different network layer combinations of the same model. In the process of image analysis, the gradient change process of Tensorboard visual loss function is used in the analysis process.3. based on Node. JS builds the server end program of image rendering style of image ink painting, and uses Express as the Web end frame, and the page effect of the front end is finished through HTML+CSS+JavaScript.
【学位授予单位】：南京大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP18;TP391.41

【相似文献】