Research On Complex Facial Expression Generation Based On Deep Learning

The generation of complex facial expression images is a research hotspot in deep learning.In the virtual battlefield deduction,multi-agents recognize the enemy's complex expressions and make strategic adjustments based on the different expressions reflected on the face.The generation of high-quality facial expression images can improve the accuracy of multi-agent expression recognition,and prompt the deduction system to efficiently make optimal combat strategies.Police officers learn the winning strategies to be performed in different situations,stop crimes on the spot in a timely and accurate manner,and provide experience in dealing with complex real scenes,which has important practical value for ensuring the safety of the country and the people.The essence of complex facial expression generation is that the expression intensity is controllable,but the current problem is that the generated image quality is not good,artifacts are easy to produce in the facial muscle area with dense expression,and the target expression intensity is weak.The research content of deep learning is diverse and has a lot of room for development.Therefore,the method of deep learning is chosen to solve the current image quality generation problem.In this paper,further research on the subject of complex facial expression generation based on deep learning,the specific work is as follows:(1)This paper introduces the current research status of generating complex facial expressions at home and abroad,expounds two classic generative confrontation network derivation models,and experimentally verifies their ability to generate complex facial expressions.Lay a research foundation for constructing MFF-GAN,a complex facial expression generation model based on deep learning.(2)Analyzed the existing facial expression image data set types and AU label coding system,and performed face alignment and cropping and AU label production preprocessing on the selected data set to provide data conditions for generating high-quality and fine-grained facial expression images.The AU label and head pose distribution of the preprocessed image are statistically analyzed,and a self-made data set suitable for MFF-GAN model training is made.(3)A model MFF-GAN is constructed to realize complex facial expression generation,which is composed of two generators and a discriminator.The encoding and decoding times of the generator keep the same data dimension.The multi-scale feature fusion module MFF fuses and modifies the coding features from the current coding layer and the next low resolution,and improves the quality of image generation after expression changes.A loss consistency loss is set between the input and reconstructed original expression images to ensure that the identity information of the characters remains unchanged during the process of expression changes.(4)The MFF-GAN model uses the training strategy of the generator and the discriminator to update the network weights,and adjusts the network training parameters in real time according to the trend of the loss function curve during training to improve the generation performance of the network.Under the Pytorch framework,the MFF-GAN model was trained and tested with a self-made complex expression data set,and compared with multiple advanced models.Among them,the comparison result with the best-performingGANimation model shows that the quality of the generated facial image is improved by 2.52,and the accuracy of AU generation accounting for 88.23% is higher than that ofGANimation.At the same time,it also qualitatively verifies that the facial expression control ability of the MFF-GAN model is enhanced when there are no blurs and artifacts in the generated image.