Scholarlink

<p>In recent years, Generative Adversarial Networks (GANs) in the field of image super-resolution have faced challenges where the local receptive field of traditional convolutions struggles to capture long-range dependencies between pixels, and implicit adversarial loss fails to accurately distinguish between real details and artifacts. These issues result in reconstructed images suffering from a lack of high-frequency details and high computational costs. To address this, this paper proposes RGSRGAN, a novel generative adversarial network (GAN) method based on recursive gated convolution for image super-resolution. First, the gated convolution and recursive design-based recursive gated convolution (g<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11760_2025_4711_Article_IEq1.gif" Format="GIF" Height="8" Rendition="HTML" Resolution="72" Type="Linedraw" Width="10" /> </InlineMediaObject> <EquationSource Format="TEX">\(^n\)</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mrow /> <mi>n</mi> </mmultiscripts> </math></EquationSource> </InlineEquation>Conv) aimed at achieving high-order spatial interaction is incorporated with the Global Filter (GF) to construct Recursive Gated Network Modules (RGNM), which are employed in the generator network to enable more effective extraction of high-order features and local details. Second, a Feature Optimization Module (FOM), constructed using recursive gated convolution and channel attention (CA), is employed to capture channel-level information and learn inter-channel dependencies, optimizing feature extraction. Finally, through the local discriminative learning (LDL) method, an artifact discriminative loss function is introduced to enable the generative adversarial network to stably generate perceptually realistic details while suppressing visual artifacts. The experimental results show that RGSRGAN recovers more high-frequency texture details, achieving superior reconstruction performance and visual quality while significantly reducing training costs.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

RGSRGAN: Image Super-Resolution Reconstruction Using Generative Adversarial Networks Based on Recursive Gated Convolution

  • 期刊论文

  • 发表日期：

    2025年9月11日

查看全文
  • Lingguang Kong,
  • Han Zhong

摘要

In recent years, Generative Adversarial Networks (GANs) in the field of image super-resolution have faced challenges where the local receptive field of traditional convolutions struggles to capture long-range dependencies between pixels, and implicit adversarial loss fails to accurately distinguish between real details and artifacts. These issues result in reconstructed images suffering from a lack of high-frequency details and high computational costs. To address this, this paper proposes RGSRGAN, a novel generative adversarial network (GAN) method based on recursive gated convolution for image super-resolution. First, the gated convolution and recursive design-based recursive gated convolution (g \(^n\) n Conv) aimed at achieving high-order spatial interaction is incorporated with the Global Filter (GF) to construct Recursive Gated Network Modules (RGNM), which are employed in the generator network to enable more effective extraction of high-order features and local details. Second, a Feature Optimization Module (FOM), constructed using recursive gated convolution and channel attention (CA), is employed to capture channel-level information and learn inter-channel dependencies, optimizing feature extraction. Finally, through the local discriminative learning (LDL) method, an artifact discriminative loss function is introduced to enable the generative adversarial network to stably generate perceptually realistic details while suppressing visual artifacts. The experimental results show that RGSRGAN recovers more high-frequency texture details, achieving superior reconstruction performance and visual quality while significantly reducing training costs.