With the widespread use of digital images in medical diagnosis, security monitoring, and entertainment media, the demand for high-resolution images has grown significantly. However, low-resolution images remain prevalent due to limitations in acquisition devices, transmission bandwidth, and storage conditions. Therefore, efficiently and accurately reconstructing low-resolution images into high-resolution images is not only theoretically important but also highly valuable in practical applications. This project focuses on super-resolution reconstruction and proposes an end-to-end mapping function based on generative adversarial networks (GANs). By leveraging the adversarial structure between the generator and discriminator, the generator learns to recover high-resolution images with rich details and textures from low-resolution inputs, thereby significantly improving image quality. The generator uses a residual network as its backbone, incorporating 16 residual blocks for multi-level feature extraction and sub-pixel convolutional layers for upsampling. The discriminator, based on a multi-layer convolutional structure, learns high-frequency details through progressive downsampling and channel expansion to distinguish between generated and real images. The experimental results show that the proposed method achieves a peak signal-to-noise ratio (PSNR) of 32.5 dB and a structural similarity index (SSIM) of 0.92 for 4× super-resolution reconstruction. The processing time for a single 512×512 image is less than 0.5 seconds with GPU acceleration, meeting real-time requirements and demonstrating strong potential for practical applications.