Face image super-resolution reconstruction based on an improved Real-ESRGAN

Face image super-resolution reconstruction aims to restore high-resolution details from low-resolution facial images and is widely applied in public security, video surveillance, and multimedia processing. To address the deficiencies of traditional methods in high-frequency texture restoration and structural consistency, this study adopts Real-ESRGAN as the baseline and integrates a Texture Enhancement Module (TEM) to enhance high-frequency texture details, a Convolutional Block Attention Module (CBAM) to strengthen key feature representations, and a Perceptual Color and Structure Loss (PCSL) to maintain color and structural stability. Experiments conducted on the CelebA-HQ dataset demonstrate that the proposed method achieves a 0.90 dB improvement in PSNR and an SSIM of 0.8296 over the baseline model, with more natural and refined subjective visual quality, providing an effective solution for high-quality face super-resolution reconstruction.