Face Image Inpainting Method Based on Multi-Scale Feature Fusion

Abstract

When dealing with severely corrupted images,the traditional image inpainting methods often produce blurry or too smooth regions in the restored images,and have difficulty in reconstructing a reasonable face image structure.To address the problem,this paper introduces a multi-scale feature fusion method into the discriminator of the traditional Generative Adversarial Network(GAN),which directly adds the upsampled feature maps of different depth to achieve effective fusion of shallow and deep information.Then this method grasps the overall pattern of the image with the help of high-level features,and fills the detail texture of the face image with low-level features,so that the resolution of the image can be fused with its semantic features for effective face image restoration.The experimental results on the dataset of CelebA show that the proposed method outperforms the regional normalization method in terms of Peak Signal to Noise Ratio(PSNR),structural similarity,L1 loss indicators,achieving ideal visual effects.