Research On Face Image Synthesis Based On Multi-task Enhanced Generative Adversarial Networks

The study of human faces in the present field of computer vision helps computers comprehend human emotions so that they can better serve humans.Face-oriented image synthesis tasks refer to the transformation of target attributes while preserving other facial features.However,due to the peculiarity and complexity of human faces,generating photorealistic images is a big challenge for this task.Traditional image synthesis methods lack pertinence because to the uniqueness and complexity of faces,while face image synthesis algorithms based on deep learning often utilize a one-to-one mapping network,resulting in low synthesis efficiency.This paper investigates the image synthesis of multitask enhanced generative adversarial networks,constructs a multi-task learning framework,and proposes a dual-domain generative adversarial network based on attention mechanism and a hierarchical adversarial network,all based on existing research on face image synthesis algorithms.Residual mask improves the quality and stability of synthetic face pictures by optimizing the network and doing training and testing on two different face datasets.The algorithm’s efficacy is confirmed by qualitative and quantitative analysis and review.Its main contents are as follows:(1)The basic knowledge of image translation and face image synthesis is briefly introduced,and several representative image synthesis algorithms are studied.On this basis,the face image synthesis network is studied.(2)The impact of extracting frequency domain feature information on picture synthesis is proven,and a dual-domain generative adversarial network based on the attention mechanism is developed.In order to address the situation where existing methods only focus on the spatial domain information,the dual-domain convolution module in the network structure simultaneously extracts the spatial domain information and the frequency domain information of the image for processing,preserving the advantages of the image in two different dimensions of the spatial domain and the frequency domain.A multi-scale attention mechanism is also included in the dual-domain generative adversarial network to make maximum use of the collected feature information.The multi-scale convolution operator,for example,naturally learns multi-scale features from coarse to fine-grained and produces a richer feature information map as an output.Finally,comparison tests are conducted on the Celeb A dataset,demonstrating the network’s efficacy and practicality in the facial image synthesis job.This approach,in comparison to the baseline methods stated in the study,makes the most of the information in the original photos and enhances the diversity of the synthesized images.(3)In order to further improve the quality of synthetic images,a hierarchical residual mask optimization network is proposed.Aiming at the particularity and complexity of human faces,this paper uses the mask image strategy to calculate the residuals between feature maps at different levels during network training as mask information,and gradually injects them into the network in a long connection manner.The calculated multi-layer mask features are fused together to output the final composite image,so as to achieve the purpose of retaining the original structural information of the character’s face.In this study,experiments were conducted on the Celeb A dataset and the Ra FD dataset,and the results proved that this method is very effective in maintaining the consistency of characters.The two image synthesis algorithms proposed in this study outperform existing models on several metrics.It provides new ideas for future face image synthesis research.