Optimized Based on Dexined: Edge-Connect’S Two-Stage Face Image Restoration

Image restoration is an important research direction in the field of computer vision. Especially in the task of face image restoration, how to generate structurally complete and visually realistic image content in the missing area has always been a challenging problem. Aiming at the problem of low edge guidance quality in traditional methods, this paper proposes a face image restoration method based on a two-stage architecture. In the first stage, the deep edge detection network DexiNed is used to replace Canny operator, and the masked image is used as input. The edge map generated by the complete image is used as the supervision signal to train a model with complete edge prediction ability. In the second stage, under the guidance of the predicted complete edge map, the generative adversarial network (GAN) is used to complete the content completion and detail restoration of the final image. Experimental results on the CelebA dataset show that the proposed method is superior to existing methods such as Edge-Connect in both PSNR and SSIM mainstream evaluation indicators. The PSNR improvement can reach up to 1.2dB, and the SSIM index is improved by about 0.14%, which is superior in terms of structure restoration and visual consistency.