结合多尺度特征和多维注意力的人脸风格转换

The Face Style Conversion Combining Multiscale Feature Fusion and Multi-Dimensional Attention

针对StarGANv2模型生成的人脸图像存在风格重建效果不佳、人脸纹理不够自然等现象,该文提出结合多尺度特征和多维注意力的人脸风格转换模型.1)将多尺度特征融合模块PSConv嵌入StarGANv2生成器内,提高了模型对图像特征的提取能力;2)提出了多维注意力模块MDConv,并将该模块嵌入StarGANv2判别器内,从而提高了模型对真假人脸图像的判别能力.与StarGANv2方法在CelebA-HQ数据集上进行对比实验的结果表明:该方法生成的人脸图像风格更美观,纹理细节更自然,学习感知图像相似度(LPIPS)的值也得到了提升.

Absrtact:According to the phenomenon that the face images generated by StarGANv2 exist poor style reconstruction effect and unnatural face texture details,a face style conversion model called MFMA-StaGANv2(multiscale feature and multi-dimensional attention StarGANv2)combining multiscale features and multi-dimensional attention is proposed.In order to improve the ability of the model to extract image features,a multiscale feature fusion module is embedded in the generator of StarGANv2.In order to improve the ability of the model to distinguish true and false face images,a multi-dimensional attention module called MDConv is proposed and embed in the StarGANv2 discriminator.Compared with StarGANv2 on CelebA-HQ dataset,the results show that the style of the face images generated by our method is more beautiful,the details of face texture are more natural,and the values of LPIPS(learned perceptual image patch similarity)are improved.