Fig. 1. Structure of multi-layer progressive guidance network model

Fig. 2. Structure of discriminator module

Fig. 3. Structure of feature extraction module

Fig. 4. Structure of FEAM module

Fig. 5. Structure of CAM

Fig. 6. Examples of different mask types and face-images. (a1)‒(a3) Mask images with missing region ratio of (0, 0.2]、(0.2, 0.4]、(0.4, 0.6]; (b1)‒(b3) face images in the CelebA-HQ dataset; (c1)‒(c3) face-images in the profile dataset

Fig. 7. Training process of the profile dataset

Fig. 8. Inpainting results of images with different mask rates in the CelebA-HQ dataset by different models. (a) Mask images; (b) EC^[16]; (c) SF^[17]; (d) PGD-Net^[15]; (e) PDGG-Net^[32]; (f) proposed model; (g) real images

Fig. 9. Inpainting results by different fine-tuning methods on the profile dataset. (a) Mask images; (b) directly load pre-training weights of CelebA-HQ dataset; (c) train directly with the profile dataset; (d) multi-dataset joint training and fine-tuning; (e) real images

Fig. 10. Inpainting results by different models on the profile dataset. (a) Mask images; (b) PGD-Net^[15]; (c) PDGG-Net^[32]; (d) proposed model; (e) real images

Fig. 11. Inpainting results of ablation experiments on the CelebA-HQ dataset. (a) Mask images; (b) without FEAM module; (c) without structural branch; (d) without texture branch; (e) proposed model; (f) real images

Fig. 12. Inpainting results of ablation experiments for the CAM on CelebA-HQ dataset. (a) Mask images; (b) without CAM; (c) without the FEAM module in CAM; (d) proposed model; (e) real images