Research On Face Frontalization Algorithm Based On Generative Adversarial Network

Nowadays,facial recognition technology has made significant progress,and its recognition performance on multiple facial datasets has approached human-level accuracy.However,these achievements are limited to frontal face recognition.Studies have shown that the accuracy of the best facial recognition technology decreases significantly for non-frontal face recognition.Facial frontalization aims to correct the profile face to a standard frontal face as a preprocessing method to improve facial recognition accuracy without changing existing facial recognition models.With the proposal of generative adversarial networks,the realism of2 D image generation has been greatly improved,so many researchers have carried out facial frontalization work based on generative adversarial networks.Existing facial frontalization methods still face the following problems during face correction:(1)Existing methods focus too much on training data,and the models trained on one dataset have reduced performance when tested on other datasets,indicating overfitting;(2)Existing methods mix all angle face data in one network,resulting in poor generation effects for different angles.Therefore,this paper proposes the following main works:(1)Propose a facial frontalization generative adversarial network based on auxiliary tasks and Transformer.To solve the problem of overfitting caused by existing methods,this method uses the correlation of multiple tasks to improve the effectiveness of facial frontalization and model generalization.The primary task is facial frontalization itself,using side-face to generate corresponding frontal-face.The secondary task is side-face sketch generation for corresponding frontal-face sketch,which assists and guides the primary task and accelerates network convergence.The two tasks share network weights and use feature interaction modules based on visual Transformer for sufficient interaction between the two parts of features.Qualitative and quantitative experiments show that this method helps to alleviate overfitting problems.(2)Propose a facial frontalization generative adversarial network based on posture attention for specific angle facial frontalization.To solve the problem of poor generation effects for different angles in existing methods,this method consists of multiple sub-networks,each trained on specific angle data.Then,using a class ensemble learning method to combine multiple sub-networks to make them applicable to any angle of facial input.Feature visualization technology is used to demonstrate the attention areas of each layer of the network,verifying the effectiveness of this method and the rationality of each module design.(3)Design and develop a facial frontalization system using Py Qt.The main functions include face detection,face alignment,image cropping,and facial frontalization based on the methods proposed in this paper.The various functional modules work together to complete the facial frontalization task,and the main interface of the system is displayed in the paper.