Face Photo-Sketch Synthesis Based On Conditional Adversarial Networks

The human face sketch hasing a rich shadow texture and a strong three-dimensional texture,can vividly express the personality characteristics of the person,and is widely used in the field of legal implementation and life.This paper proposes a sketch generation method based on sketch order constraint,about how to use the intermediate results generated by the painter in the sketch drawing process to fully constrain the output of different intermediate layers of the depth network for improving the realisticity of the final sketch generation method.The main work of the article has the following three aspects:(1)firstly,build a database containing 400 face sketches,which we call Ord-Sketch(order-Sketch with sketch order).In the database,each face correspondence Includes a 25-step sketch sequence diagram.The database is drawn according to the CUHK student database and the CUSFS non-Chinese identity database;(2)Ord-Sketch is a database with time series characteristics.According to it,we proposed a SO-Net(Self-Organizing Network)model based on the generated conditional confrontation network.The SO-Net model divides the sketch generation process into five phases(Since we only choose 5 intermediate results from the 25-step intermediate results from the Ord-Sketch database,which the painter believes are the most representative results).The SO-Net model has five generators and five discriminators,and the generator uses the Convolutional Long Short-Term Memory(ConvLSTM),which has unique advantages in processing sequence problems,to learn the timing of the sketch sequence information.Each stage generator contains a Convolutional Neural Network(CNN)encoder and decoder of the same structure.Unlike with other CNN methods,before the output of Feature information,each stage encoder is first enhanced by ConvLSTM.Finally,the face is sketched by the decoder.This design simulates the human sketching process,that is learning the contours first,and then gradually describing the details,allowing the network to gradually learn more features.The discriminator adopts PatchGAN to make the generated sketch with high resolution and maintain the image detail.(3)Based on the SO-Net network framework,this paper designs a loss function based on the sketch-order sketch generation.method,which effectively improves the quality of the generated sketch.Finally,the experimental results on the Ord-Sketch database test set proves that the sketch generation method based on the sketch order constraint can improve the quality of sketch generation compared with other methods in the literature.At the same time,we have built a database of 25-step sketch(or intermediate results),which will also provide some reference and help for other researchers.