计算机工程

• •    

基于多尺度特征融合的人脸图像修复方法

  

  • 发布日期:2020-12-17

Face image inpainting method based on multi-scale feature fusion

  • Published:2020-12-17

摘要: 为了解决传统方法在修复受损区域较大的图像时,产生的结果过于平滑或模糊且难以重建合理结构的问题,本文提 出在传统的生成对抗网络的鉴别器中引入多尺度特征融合的方法,将不同深度的特征图经过上采样后直接相加,有效地融合 了浅层以及深层的信息。该方法本质上是借助高层特征把握图像的整体规律,同时利用低层特征填充图像的细节纹理,进而 可以有效地将一张图像的分辨率及其语义特征相互融合。实验结果表明,在 CelebA 数据集上,本文提出的多尺度特征融合的 人脸修复方法在峰值信噪比、相似性结构、L1 损失等测度上达到了最佳效果,同时也取得了理想的视觉效果。

Abstract: In order to solve the traditional method in the repair the damaged area larger images, the result of too smooth or vague and difficult to rebuild the reasonable structure of the problem, this paper puts forward a multi-scale feature fusion method in the discriminator of the traditional Generative Adversarial Networks (GANs),the introduction of the characteristics of different depth chart after sampling on direct addition, combined with the shallow and deep information effectively. In essence, this method grasps the overall law of the image with the help of high level features and fills the detail texture of the image with low level features, so that the resolution of an image and its semantic features can be effectively fused with each other. The experimental results show that the multi-scale feature fusion method proposed in this paper achieves the best results in terms of Peak Signal to Noise Ratio , Structural Similarity, L1 loss and other measures on CelebA data set, and also achieves ideal visual effects.