基于拉普拉斯算子抑制伪影的神经风格迁移方法  

Neural Style Transfer Method Based on Laplace Operator to Suppress Artifacts

作　　者：张美玉[1] 刘跃辉 秦绪佳[1] 吴良武 ZHANG Mei-yu;LIU Yue-hui;QIN Xu-jia;WU Liang-wu(College of Computer Science and Technology,Zhejiang University of Technology,Hangzhou 310023,China;Dalian Institute of Test and Control Technology,Dalian,Liaoning 116013,China)

机构地区：[1]浙江工业大学计算机科学与技术学院,杭州310023 [2]大连测控技术研究所,辽宁大连116013

出　　处：《计算机科学》2020年第S02期209-214,共6页Computer Science

基　　金：国家自然科学基金(61672463);浙江省自然科学基金(LY20F020025,LY18F020035)。

摘　　要：在图像神经风格迁移(Neural Style Transfer)技术中,大多算法都存在影响视觉效果的伪影:棋盘效应与影响原图语义内容的纹理。对此,提出一种基于拉普拉斯算子抑制伪影的图像风格迁移方法。首先,使用空洞卷积、1×1卷积重新设计了快速神经风格迁移的转换网络。然后,将变换后的结果输入VGG进行特征检测,并将原图也输入VGG进行特征检测,将这两种特征进行拉普拉斯算子滤波后计算两者的L1误差。约束图像变化,以抑制伪影。在最后的解码器阶段,使用了重新设计的网络结构,并增加了dropout的编码器来修改图像内容。在加深网络的同时,通过1×1卷积控制模型体积,将模型体积缩小了6%。实验表明了该方法抑制伪影的效果优于传统方法,其可以生成良好视觉效果的图像。In image neural style transfer technology,most algorithms have artifacts that affect visual effects:checkerboard effects and textures that affect the semantic content of the original image.In this paper,an image style transfer method based on Laplacian suppression artifacts is proposed.Firstly,a transformation network for real-time neural style transfer is redesigned using hole convolution and 1×1 convolution filter kernels.Then,the transformed result is input to VGG for feature map detection,and the multi-layer feature and the original VGG feature are extracted and filtered by Laplace operator to calculate the L1 error.Constrain image changes to suppress artifacts.In the final encoder stage,the image content is modified using an encoder with added dropout.While deepening the network,the model size was controlled by 1×1 convolution filter kernels,which reduced the model size about 6%.Finally,experiments show that the results of this method are better than traditional methods in suppressing artifacts,and can produce images with good visual effects.

关 键 词：风格迁移 卷积神经网络 拉普拉斯算子 GRAM矩阵 残差 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]