A Study on an Improved Generative Adversarial Network Model for Super-Resolution Image Reconstruction

The SRGAN model based on Generative Adversarial Networks (GANs) has achieved breakthroughs in perceptual quality for image super-resolution reconstruction. However, existing models suffer from feature smoothing effects in the generator and unstable training of the discriminator. To address these limitations, we propose an improved SRGAN model (denoted as SRGAN-GBDB) through multi-strategy collaborative optimization to improve reconstruction quality. The specific improvements include: Firstly, in the generator, we remove batch normalization (BN) layers to alleviate feature degradation and embed the Convolutional Block Attention Module (CBAM) to enhance high-frequency feature modeling. Secondly, we implement spectral normalization for the discriminator and incorporate spectrally normalized residual blocks into the discriminator to strengthen its discriminative capability. Additionally, the joint perceptual-adversarial-content loss function of the SRGAN model balances reconstruction accuracy and visual quality. Experimental results show that the optimized model SRGAN-GBDB outperforms the traditional Bicubic interpolation method, SRGAN, and several other existing super-resolution algorithms on the Set5, Set14, BSD100, Urban100, and DIV2K datasets. It not only improves quantitative metrics such as PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index) but also demonstrates superior visual quality through better detail recovery and fewer artifacts, validating the effectiveness and superiority of the optimization strategy. Future work will explore more efficient optimization methods and lightweight network architectures to facilitate the application of image super-resolution systems in mobile platforms and specialized fields like remote sensing and medical imaging.