PKDFIN: Prior Knowledge Distillation-Based Face Image Inpainting Network for Missing Regions

Abstract

Existing facial image inpainting methods demonstrate high reliance on the precision of prior knowledge. However, the acquisition of precise prior knowledge remains challenging, and the incorporation of predicted prior knowledge in the restoration process often leads to error propagation and accumulation, thereby compromising the reconstruction quality. To address this limitation, we propose a novel facial image inpainting framework that leverages knowledge distillation, which is specifically designed to mitigate error propagation caused by imprecise prior knowledge. More specifically, we develop a teacher network incorporating accurate facial prior information and establish a knowledge transfer mechanism between the teacher and student networks via knowledge distillation. During the training phase, the student network progressively acquires the prior information encoded in the teacher network, thus improving its restoration capability for missing or corrupted regions. Additionally, we introduce a Coordinate Attention Gated Convolution (CAG) module, which enables effective extraction of both structural and semantic features from intact regions. Experiments conducted on the public facial datasets (CelebA-HQ and FFHQ) show that our method achieves performance improvements over existing approaches in terms of multiple quantitative evaluation metrics, including PSNR, SSIM, MAE, and LPIPS. Thus, the knowledge transfer from teacher to student network via knowledge distillation significantly reduces the dependence on prior knowledge characteristic of existing methods, facilitating more precise and efficient facial image inpainting.