基于深度残差生成对抗网络的超分辨率重建算法研究

Research on Super-Resolution Reconstruction Algorithms Based on Deep Residual Generative Adversarial Networks

摘要: 图像超分辨重建是计算机视觉领域中的重要研究方向之一。本文主要针对图像超分辨率在重建过程中信息恢复不充分、特征提取不全面、高频细节不明显等问题，在SRGAN的基础上提出一种基于深度残差生成对抗网络的图像超分辨率算法。该算法实现了一种深度残差的结构(Depth-ResNet)，即残差中的残差，来形成非常深的网络。该结构由几个具有长跳跃连接的残差组组成，每个残差组中也包含一些具有短跳跃连接的残差块。除此之外，Depth-ResNet允许通过多个跳跃连接绕过丰富的低频信息，使主网络专注于学习高频信息，并且随着Depth-ResNet的数量与深度的调整变化，图像重建效果将取得更好的准确性与视觉改进。此外，为了平衡Depth-ResNet在生成器上的卓越性能，本文在生成器的损失函数上采用了Charbonnier损失函数与对抗损失函数，并优化了判别器的结构。根据大量实验表明，重建的图像在清晰度、高频细节等方面都有一定的提高。

Abstract: Image super-resolution reconstruction is one of the significant research directions within the field of computer vision. This paper particularly addresses issues such as insufficient information recovery, incomplete feature extraction, and indistinct high-frequency details during the image super-resolution process by proposing, on the basis of SRGAN, a novel deep residual generative adversarial network-based image super-resolution algorithm. The proposed algorithm implements a Depth-Residual Network (Depth-ResNet) architecture, characterized by residuals within residuals, to construct an extremely deep network. This architecture comprises several residual groups interconnected with long skip connections, where each residual group further consists of several residual blocks linked via short skip connections. Significantly, the Depth-ResNet design allows for multiple skip connections bypassing rich low-frequency information, enabling the main network to focus on learning high-frequency information effectively. As the number and depth of the Depth-ResNet modules are varied and adjusted, the reconstructed images exhibit improved accuracy and visually enhanced outcomes. Moreover, in order to balance the outstanding performance of the Depth-ResNet in the generator, this study adopts the Charbonnier loss function along with the adversarial loss function in the generator’s objective function, while also optimizing the structure of the discriminator. Extensive experimentation demonstrates that the reconstructed images exhibit noticeable improvements in terms of clarity and high-frequency detail preservation.