SUPERGAN:Facial Prior-based Landmark Localization Network For Super-Resolution Arbitrary Image Attribute Editing

Incorporated encoder-decoder and generative adversarial networks are a popular solution for arbitrary attribute editing task.Nevertheless,blurry editing and low quality result usually is obtained because of the bottleneck layer in encoder-decoder networks.The result image quality can be improved by adding skip connections,but only at the expense of loosen attribute manipulation capability.Furthermore,generally target attribute vector in existing methods is used to conduct the supple translation to the acceptable target domain.In our work,we show that selective transfer perspective can be exploit to solve these issues.We present an original trainable deep Facial Prior-based Landmark Localization Network,which uses the geometry prior,that consisted of facial parsing maps and landmark heatmaps,combined in Facial Prior Evaluation Network.Experiments show that using our approach,we can get super-resolution face images from the very low-resolution face images with the scale factor 8,even if those input face images are not well aligned for the Arbitrary Attribute Editing task.In our model we consider that the certain editing assignment instead of all target attributes is clearly only related to the transformed attributes.Thus,encoder-decoder is incorporated with selective transfer units and prior evaluation network to arbitrary select and define encoder feature for the amplifying attribute editing ability and obtain super resolution results.Analyzing of various experiments results in our work presents that our method,called Super GAN,improves attribute generation accuracy,even with using very low-resolution face images,and,at the same time,increase the face resulting images perception quality.Moreover,in this paper we compare our approach with state-of-the-arts methods in the arbitrary face attribute editing field,and various experiments shows that our Super GAN brings better results.In other words,to generate not only realistic face images,but also super-resolution realistic face images,in our work we propose Super GAN.Further,we introduce the new estimation metrics for the better face SR visual perception,related with the few super-resolution tasks: face alignment and parsing;and methods of how to extract the facial image features,which we present in our prior evaluation network.In our work we provide extensive experiments,which show that our Super GAN highly outperforms other face hallucination methods,and,moreover,brings up significant results in arbitrary attribute editing perception quality.