Year: 2024
Type: conference-paper
Abstract: To solve the structural distortion and artifacts of existing face image inpainting algorithms under large-area mask, this paper proposes a generative face image inpainting algorithm based on predictive filtering and multi-priority guidance. The inpainting process of this algorithm is divided into two stages. In the first stage, firstly, the face sketch of masked face is utilized as a prior information to guide the edge prediction network to generate the complete face edges. Then the landmark information in the masked region of the face image is predicted using Mobilenet. In the second stage, the face sketch, the predicted edges and landmarks are used as prior information, and we introduce feature-level and image-level predictive filtering in the face inpainting module to obtain the high-quality face inpainting results. The experimental results on Celeba-HQ dataset and FFHQ dataset show that the proposed algorithm can achieve better inpainting results for large-area masked face. Compared with other six representative face inpainting algorithms, the proposed algorithm achieves effective improvement in three criterion: PSNR, SSIM and FID.
Authors: Fuping Wang, Jiahao Wang, Zhangyan He
Institution: Xi’an University of Posts and Telecommunications
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4411730248