The combination of Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs) for image super-resolution reconstruction
Year: 2025
Type: article
Abstract: This study developed a hybrid model combining a Convolutional Neural Network (CNN) and a Generative Adversarial Network (GAN) for the task of single-image super-resolution reconstruction. The CNN is responsible for hierarchical image feature extraction and maintaining structural integrity, while the GAN synthesizes realistic texture details through an adversarial training mechanism to enhance visual realism. The generator is constructed using densely connected convolutional blocks and is combined with an image block-based discriminator to evaluate the authenticity of local regions. The composite loss function is designed to integrate the root mean square error, perceptual loss, and adversarial loss of the pre-trained GTS network, balancing pixel-level accuracy and visual perceptual effect. Tests on benchmark datasets such as DIV2K and Set14 show that this model outperforms traditional interpolation algorithms and deep learning models in objective indicators such as PSNR and SSIM, as well as in the perception evaluation of LPIPS. Especially in complex texture restoration tasks, the model demonstrates excellent detail restoration capabilities. Experimental data confirm that the adversarial training mechanism effectively solves the common problem of excessive smoothing in traditional super-resolution methods, making the reconstructed image closer to the actual optical imaging effect. This technology provides new ideas for scenarios that require high-fidelity reconstruction, such as medical image analysis and satellite map optimization.
Author: Kaitian Chai
Institution: Australian National University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4411041717