Research On Generative Adversarial Network Of Facial Image Generation Based On Self-encoder Structure

Facial image is a common data in daily life and it plays an important role in various application situation.With the continuous development and breakthrough of Generative Adversarial Network in recent years,the effect of facial generated image technology in the field of facial image processing is more and more obvious.It can not only be directly applied in people's daily life,but also assist in facial image analysis technology,providing data and technical support.Although facial generated image technology is becoming more and more perfect and the effect is amazing.It can not only generate realistic facial image with high quality,but also generate facial image with high resolution.However,there are still some problems,such as poor controllability,the complex of model structure and large amount of parameters and so on.The paper design a generative adversarial network based on the structure of an autoencoder,which can not only generate facial images,but also generate encoding vectors of facial images for facial image reconstruction and control of facial generated image.The model introduces two pearson correlation coefficients,one is used to complete the facial image encoding and the other is used to stabilize the quality of the facial image.At the same time,in order to balance the image encoding and image generation,the model adds adaptively trained dynamic transformation coefficients to adjust the weight ratio of the two pearson correlation coefficients in time according to the training situation.The designed model does not require additional discriminators.The model is simple and the parameters are small.The experimental results show that the proposed generative adversarial network achieves both facial image encoding and facial image generation.In view of the poor quality and poor diversity of facial images generated by small models,the paper introduces Self-mod and NetVLAD modules.Self-mod is based on the conditional BN,and uses the random noise vector as the input of the conditional BN.It obtains two learnable parameters that are dependent on the input random noise vector through training,and participates in the adjustment of the BN layer.NetVLAD is based on the VLAD vector.By softly assigning coefficients,the VLAD vector can be used as the network layer of the model for back propagation of parameters.NetVLAD can well reflect the similarity between images.Experiments show that after adding two modules,the quality and diversity of the generated face images have improved,and FID is better.In summary,the model designed in this paper can generate better face images and coding vectors with a smaller amount of parameters,and solves the problem of poor quality and mode collapse of face images generated on small models.