Design of Face Image Super-Resolution Network Based on Multi-scale Perceptive Inception Dense Block

Within the realm of image data manipulation, enhancing resolution and reconstructing facial visuals stand as pivotal methods designed to restore poorly defined images impacted by unidentified deterioration mechanisms. Traditional non-generative adversarial network (non-GAN) models suffer from high-frequency information reconstruction deficiencies when restoring high-frequency information in images. Utilizing the adversarial learning process between generators and discriminators, GANs are capable of synthesizing high-definition imagery enriched with authentic texture and enhanced clarity. However, they have inherent limitations in terms of pixel-level changes, line distortions and detail recovery. In the context of Mixed Residual of Multi-scale Perceptive Inception Dense Block GAN (MRMPIDBGAN), a Multi-scale Perceptive Inception Block (MPIB) is engineered for the robust retrieval of features and multi-scale data from facial images. Based on MPIB, a Mixed Residual of Multi-scale Perceptive Inception Dense Block (MRMPIDB) is further constructed to form the generator, which facilitates better capture and reconstruction of facial image details. Secondly, to more accurately capture the local texture information of the image, the original VGG discriminator in ESRGAN is replaced by a U-Net discriminator, and spectral normalization techniques are applied to improve the stability of the discriminator. Subsequently, an innovative upsampling technique is formulated to optimize efficacy yet minimize computational demands. Further improvements in the model's efficacy are achieved through the replacement of the traditional VGG loss by a metric grounded in perceptual criteria. Experimental results confirm the effectiveness of this algorithm in super-resolution facial image restoration and demonstrate significant advantages in facial image reconstruction, clear texture capture and feature detailing compared to existing methods.