【作者】 王薇；

【导师】 李毅； 谭诗瀚；

【作者基本信息】 四川大学 ， 工程硕士（专业学位）， 2025， 硕士

【摘要】 表情作为人类表达内心情感的重要载体,在人际交往和社会互动中扮演着关键角色。通过面部表情的细微变化,人们能够准确感知他人的情绪状态,实现非语言的情感交流。表情生成技术旨在保持个体身份特征不变的前提下,将给定的面部表情转换为目标表情,这一技术在影视剧拍摄与制作、摄影制作的后期处理、虚拟现实等各行各业领域都具有广泛的应用前景。近年来,随着计算机视觉和深度学习等互联网技术的快速发展,面部表情生成方向的研究已成为学术界关注的热点课题。生成对抗网络(Generative Adversarial Network,GAN)在图像生成领域有着极为广泛的应用,凭借其强大的生成能力,在面部表情生成领域取得了显著进展,然而,当前方法仍然存在一些局限性。一是生成面部表情的逼真程度和对齐程度不足,在表情变化剧烈的区域常出现细节失真和边界模糊现象,二是GAN的训练不够稳定,存在梯度消失等问题,针对以上问题,本文从提高生成图片质量和改进GAN的训练两方面提出了解决方案。本文工作主要包括以下几个方面:一、在图像质量提升方面,本文提出了一种基于特征空间优化的生成对抗网络(Feature Space Optimization Generative Adversarial Network,FSO-GAN)。该模块的核心思想是通过优化特征空间分布来提升生成效果。具体而言,该方法首先构建困难样本和易样本的分类体系,其中困难样本指代那些难以被分类器准确识别的图像数据。通过线性组合真实样本的方式构造困难样本,同时在生成图像中引入可控噪声以增强模型的鲁棒性。在此基础上,设计了一个高效的特征提取器,通过最小化困难样本与易样本在特征空间中的距离,有效压缩了判别器特征空间中训练数据的分布范围,从而显著提升了生成图像的真实感和细节表现力。二、在稳定模型训练方面,本文提出了一种基于稳定梯度的生成对抗网络(Stable Gradient-based Generative Adversarial Network,SGrad GAN)。由于GAN在训练初期容易过早进入理想状态,导致生成器无法获得有效的梯度反馈,参数更新停滞,进而引发梯度消失和训练震荡等问题。SGrad GAN的核心思想是在判别器中引入梯度稳定算法,直接对判别器的输出进行梯度约束,将网络的参数变化范围控制在合理区间,通过这种机制不仅可以有效缓解训练过程中的模式崩溃,还可以提升模型收敛的稳定性。三、在CelebA数据集和PubFig数据集上进行了大量实验,选择了PSNR、SSIM和LPIPS三个指标进行了对比分析,可视化结果和指标分析都表明,本文提出的表情合成算法和基线模型相比,效果均有所提升,此外,还进行了消融实验和超参数分析,进一步验证了各个模块在模型中的作用,以及不同的参数对模型的影响,充分证明了所提出算法的有效性。更多还原

【Abstract】 Expression,as an important carrier for humans to express their inner emotions,plays a crucial role in interpersonal communication and social interaction.Through the subtle changes in facial expressions,people can accurately perceive the emotional states of others and achieve non-verbal emotional communication.The expression generation technology aims to transform a given facial expression into a target expression while keeping the individual identity features unchanged.This technology has broad application prospects in various fields such as film and television drama shooting and production,post-processing of photography production,and virtual reality.In recent years,with the rapid development of Internet technologies such as computer vision and deep learning,the research in the field of facial expression generation has become a hot topic in the academic community.The Generative Adversarial Network(GAN)has extremely wide applications in the field of image generation.With its powerful generation ability,it has made remarkable progress in the field of facial expression generation.However,the current methods still have some limitations.Firstly,the generated facial expressions lack in realism and alignment.There are often details distortion and boundary blurring phenomena in the areas with drastic expression changes.Secondly,the training of GAN is not stable enough,and there are problems such as gradient vanishing.In response to the above problems,this paper proposes solutions from two aspects:improving the quality of generated images and improving the training of GAN.The main work of this paper includes the following aspects:Firstly,in terms of improving image quality,this paper proposes a Feature Space Optimization Generative Adversarial Network(FSO-GAN)based on feature space optimization.The core idea of this module is to enhance the generation effect by optimizing the distribution of the feature space.Specifically,this method first constructs a classification system for hard samples and easy samples,where hard samples refer to those image data that are difficult to be accurately recognized by the classifier.Hard samples are constructed by linearly combining real samples,while controllable noise is introduced into the generated images to enhance the model’s robustness.On this basis,an efficient feature extractor is designed,which minimizes the distance between hard samples and easy samples in the feature space,effectively compressing the distribution range of training data in the discriminator’s feature space,thereby significantly improving the realism and detail performance of the generated images.Secondly,in terms of stabilizing model training,this paper proposes a Stable Gradient-based Generative Adversarial Network(SGrad GAN).Due to the tendency of GANs to prematurely reach an ideal state during the early stages of training,the generator may fail to receive effective gradient feedback,leading to stagnant parameter updates and subsequently triggering issues such as gradient vanishing and training oscillations.The core idea of SGrad GAN is to introduce a gradient stabilization algorithm into the discriminator,directly constraining the gradient of the discriminator’s output to keep the parameter variations within a reasonable range.This mechanism not only effectively mitigates mode collapse during training but also enhances the stability of model convergence.Thirdly,extensive experiments were conducted on the CelebA and PubFig datasets,with PSNR,SSIM,and LPIPS selected as evaluation metrics for comparative analysis.Both the visualization results and metric analysis demonstrate that the proposed expression synthesis algorithm outperforms baseline models.Additionally,ablation studies and hyperparameter analysis were performed to further validate the role of each module in the model and the impact of different parameters on the model,fully proving the effectiveness of the proposed algorithm.更多还原