Semantic Image Synthesis Based on Improved Cascaded Refinement Network
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 im-proved semantic image synthesis method based on improved cascaded refinement networks(CRNs)is pro-posed.In cascaded thinning network,spatially-adaptive(de)normalizationwas used instead of layer nor-malization,and activation in normalization layer was adjusted by spatial adaptive learning to make seman-tic information more complete.Smooth L1 loss function was introduced to reduce thecolor difference be-tween theoutput image and contrast image.In addition,learnable spatial adaptive normalization was intro-duced 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.
image synthesiscascaded refinement networksspatially-adaptive(de)normaliza-tionsmooth L1 loss
NETL
NSTL
万方数据
