查询结果:   王婷,李航,胡智.一种VGGNet的图像风格迁移算法设计与实现[J].计算机应用与软件,2019,36(11):224 - 228.
中文标题
一种VGGNet的图像风格迁移算法设计与实现
发表栏目
图像处理与应用
摘要点击数
308
英文标题
DESIGN AND IMPLEMENTATION OF IMAGE STYLE MIGRATION ALGORITHM BASED ON VGGNET
作 者
王婷 李航 胡智 Wang Ting Li Hang Hu Zhi
作者单位
沈阳师范大学软件学院 辽宁 沈阳 110034     
英文单位
College of Software, Shenyang Normal University, Shenyang 110034, Liaoning, China     
关键词
图像风格迁移 Keras VGGNet 深度学习
Keywords
Image style migration Keras VGGNet Deep learning
基金项目
国家自然科学基金项目(60970112)
作者资料
王婷,硕士生,主研领域:深度学习,图像识别。李航,教授。胡智,教授。 。
文章摘要
针对卷积神经网络在实现图像风格迁移中出现的图像失真及精度较差问题,提出一种基于卷积神经网络的图像风格迁移算法。分析传统的纹理重构算法,采用拟牛顿法之一的L-BFGS优化方法对其进行改进。利用Gram矩阵计算图片中的纹理、颜色和视觉信息,提取一幅普通图片和一幅具有代表性的艺术性图像的两种高层抽象特征表示,从而生成具有原内容和艺术性风格的合成图像。在深度学习Keras框架的基础上,设计一种卷积神经网络的图像风格迁移算法。实验结果表明,适度地选择迭代次数可观察合成图像的匹配程度,该算法可提高准确度并降低计算复杂度。
Abstract
Aiming at the problem of image distortion and poor accuracy in image style migration based on convolutional neural network, we proposed an image style migration algorithm based on convolution neural network. We analyzed the traditional texture reconstruction algorithm, and used the L-BFGS optimization method of quasi Newton method to improve it. Then, we used the Gram matrix to calculate the texture, color and visual information in the picture, and extracted two high-level abstract features of a common picture and a representative artistic image to generate a composite image with original content and artistic style. On the basis of Keras framework of deep learning, we designed an image style migration algorithm based on convolution neural network. The experimental results show that the appropriate number of iterations can be selected to observe the matching degree of the composite image, and the proposed algorithm can improve the accuracy and reduce computation complexity.
下载PDF全文