基于改进的MUNIT人脸图像性别转换模型

Gender Transformation Model of Face Image Based on Improved MUNIT

基于生成对抗网络的图像风格迁移算法已成为人脸图像性别转换的主流模型,但现有方法仍存在转化后的人脸图像模糊,背景图像扭曲,面部身份保留效果不好等缺点。针对上述问题,基于多模态无监督图像翻译网络(MUNIT),本文提出了基于改进的人脸图像性别转换模型。首先对MUNIT模型生成器部分进行优化,在编码器部分加入动态实例归一化操作(DIN),使编码器对人脸内容特征和风格特征的剥离更加精确;并在内容编码部分的残差块网络后加入混合注意力模块(CBAM),使模型提取更丰富的人脸关键特征;此外,对CeleBA数据集的人脸图像根据属性进行筛选以及裁剪,减少了图像背景对于生成图像的影响,使模型更加专注于人脸特征的学习。根据实验对照情况,本文方法能够生成更加精细的人脸性别转换图像。

The image style transfer methods based on generative adversarial network have become the main-stream model of face image gender transformation. However, for existing methods, the transformed face image is blurred, the background image is distorted, and the facial identity preservation effect is not good. Aiming at the above problems, this paper proposes an improved face image gender transformation model based on multi-modal unsupervised Image Translation Network (MUNIT). Firstly, the generator part of MUNIT model is optimized, and the dynamic instance normalization operation (DIN) is added to the encoder part to make the encoder more accurate in the stripping of face content features and style features. The mixed attention module (CBAM) is added after the residual block network in the content encoding part, so that the model can extract more abundant face key features. In addition, the face image of CeleBA dataset is screened and trimmed according to its attributes, which reduces the influence of image background on image generation and makes the model more focused on the learning of face features. According to the experimental situation, the proposed method can generate more refined facial gender conversion images.