Combining residual structure and edge loss for face image restoration with generative adversarial networks

Aiming at the problem of incoherence caused by the color deviation of the edges when splicing with the background after repairing the defective region of the face image, this paper proposes a generative adversarial network face image repair method with residual structure and edge loss. Firstly, a multilayer residual structure is used between the convolutional layer and the inverse convolutional layer to improve the feature extraction and characterization ability of the generator and enhance the coherence of the edges of the repaired region and the background; secondly, the edge loss is proposed to address the edge coherence problem, and the mean-square error of the outer edges of the repaired region and the inner edges of the background part is computed to be backpropagated, which further reduces the color deviation of the edges of the repaired region and the background; finally, the generator loss that adds edge loss and the discriminator loss that introduces gradient penalty are combined into a new multi-scale loss function, and different weights are set to adjust the multi-scale loss function to enhance the restoration effect to the best, accelerate the convergence speed of the model, and narrow down the difference between the restored image and the real image from the whole and the edges. In this paper, peak signal-to-noise ratio, structural similarity, and frechet inception distance are used as the evaluation criteria of the model. The experimental results show that compared with the traditional DCGAN restoration method, the coherence between the edges of the restored region and the background is significantly enhanced with a PSNR value of 30.06, which is 50.14% higher, an SSIM value of 0.935, which is 19.72% higher, an FID value of 3.25, which is 90.75% lower.