Facial Aesthetic Perception And Calculation Based On Geometric Features And Deep Learnin

Facial beauty research originates from psychology.The perception of facial beauty varies across different eras,and researchers have proposed various standards and rules throughout different periods.In the past decade,as an emerging research topic,the study of facial beauty based on computer vision has attracted more attention from scholars.Numerous studies have been conducted on facial beauty analysis and enhancement.Theoretically,using existing facial beauty datasets,the traditional methods and deep models have been developed for evaluating and enhancing facial beauty.In practical applications,beautifying softwares like Meitu Xiuxiu and Kuaishou live streaming are gaining popularity.However,facial beauty evaluation,influenced by subjective cognition and other factors,remains a challenging area of study.Given the numerous challenges in facial beauty research,this study mainly focuses on how to evaluate facial beauty and how to enhance it.Firstly,an optimized feature selection algorithm is proposed to analyze the geometric semantics of facial beauty from high-dimensional geometric features.Then,for the task of predicting facial beauty scores,combined with prior knowledge of the importance of facial geometric features,this paper designs a dual branch deep network model,with results closely aligning with human subjective evaluation scores.For face beautification tasks,this article utilizes geometric semantics in facial beauty analysis and combines generative adversarial learning theory to establish a deep learning based beautification model,which enhances facial beauty while preserving identity.The main research content includes the following sections:(1)A model for optimizing facial landmarks for beauty analysis is proposed.Influenced by traditional beauty ideas like the golden ratio,facial beauty analysis based on pattern recognition extracts geometric features through facial landmark models,subsequently using machine learning algorithms to analyze and predict facial beauty.The extracted geometric features vary with the landmark model,thus affecting beauty analysis performance.Existing studies usually rely on researchers’ subjective perceptions to select different landmark models for geometric feature extraction,resulting in lower performance in evaluating facial beauty.This study,based on public perception of facial beauty,constructs a 171-dimensional facial geometric feature set.An optimized genetic algorithm is used for feature selection,refining the existing 68 facial landmark model covering these 171 features down to a 62-landmark model,where each landmark contributes its geometric feature to facial beauty.(2)A competitive swarm optimization feature selection algorithm for facial beauty analysis is proposed.Most existing facial beauty analyses rely on empirical extraction of geometric features,often not exceeding 100 dimensions,potentially overlooking some crucial geometric features.To comprehensively study facial beauty-related geometric features and minimize subjective bias,we extract the 2928-dimensional geometric feature set including distances between all landmarks,ratios,and angles based on the aforementioned optimized 62-landmark model.To address the issue of redundant features,a dynamic weight-guided competitive swarm optimization algorithm is introduced to select important and effective geometric features for evaluating facial beauty.(3)A deep network model for evaluating facial beauty scores guided by geometric priors is proposed.As observed from the above research,some global features,such as long-distance and large-scale angles and their relationships,are crucial for facial beauty.However,existing deep network models,primarily using convolutional neural networks,struggle to capture these global geometric features adequately.This study,based on such geometric priors,designs a dual-branch network structure,employing the Swin Transformer and Resnet-50 as the backbones.The former excels in capturing global facial features,while the latter focuses on specific local features of facial components.Furthermore,a geometric regularizer based on facial landmarks is introduced to enhance the network’s learning ability on facial geometric features,thereby improving model performance.(4)A semantically interpretable face beautification model based on generative adversarial networks(GANs)is proposed.The second research further emphasizes the importance of local and global facial features in facial beauty analysis.Based on this,a new dual-branch facial beauty semantic style encoder is introduced.One branch operates on the full face,obtaining a beautiful appearance texture style code through a style encoder and semantic decomposer.The other branch operates on the facial geometric image,producing a geometric style code for beautiful faces through a mask encoder and fusion module.Additionally,face beautification is a complex image generation task.To preserve the original background information,an attention-based generator is designed.