基于改进级联细化网络的语义图像合成  

Semantic Image Synthesis Based on Improved Cascaded Refinement Network

作　　者：王佳琦 王国刚 汪滢 赵怀慈[2] WANG Jiaqi;WANG Guogang;WANG Ying;ZHAO Huaici(Shenyang University of Chemical Technology,Shenyang 110142,China;Shenyang Institute of Automation Chinese Academy of Sciences,Shenyang 110142,China)

机构地区：[1]沈阳化工大学信息工程学院,辽宁沈阳110142 [2]中国科学院沈阳自动化研究所,辽宁沈阳110016

出　　处：《沈阳化工大学学报》2024年第2期155-160,共6页Journal of Shenyang University of Chemical Technology

基　　金：国家重点研发计划项目(2018YFB1700200)。

摘　　要：针对级联细化网络(cascaded refinement networks,CRNs)存在合成图像不完整、语义信息丢失、合成的图像颜色差异大的问题,提出一种改进级联细化网络的语义图像合成方法.在级联细化网络中用空间自适应归一化[spatially-adaptive(de)normalization,SPADE]代替层归一化,通过空间自适应的学习调节归一化层中的激活,从而使语义信息更加完整;引入平滑L 1损失函数,减少输出图像和对比图像间的颜色差异;引入可学习的空间自适应归一化,增加网络参数的存储容量,能够学习更多的语义信息,使合成的图像质量得到提升.在Cityscapes数据集和GTA5数据集上的试验结果表明:该方法的平均交并比和像素准确性分别比CRNs的提升了31.4%和7.4%,弗雷歇初始距离比CRNs的降低了16.3%.In order to solve the problems in cascaded refinement networks(CRNs),such as incomplete synthetic images,loss of semantic information and large color difference of synthesized images,an improved semantic image synthesis method based on improved cascaded refinement networks(CRNs)is proposed.In cascaded thinning network,spatially-adaptive(de)normalizationwas used instead of layer normalization,and activation in normalization layer was adjusted by spatial adaptive learning to make semantic information more complete.Smooth L 1 loss function was introduced to reduce thecolor difference between theoutput image and contrast image.In addition,learnable spatial adaptive normalization was introduced to increase the storage capacity of network parameters.More semantic information could be learned to improve the quality of the synthesized image.Experiments on Cityscapes and GTA5 datasets show that the mean intersection over Union and pixel accuracy are 31.4%and 17.4%higher than thatof CRNs,respectively,and the Fréchet Inception Distance is 16.3%lower than CRNs.

关 键 词：图像合成 级联细化网络 空间自适应归一化 平滑L1损失 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]