Scholarlink

In this paper, we introduce the Dynamic Hybrid Spectral Synthesis Network (DHS2Net), a novel dual-domain learning approach. The network innovatively separates high and low frequency components using convolutional layers with different strides, subsequently directing the low-frequency elements through a sequence of residual in residual blocks for spatial domain refinement, while employing complex-valued convolutions for high-frequency feature processing. An integral component of this architecture is the Feature Enhancement and Selection Unit (FESU), which dynamically selects and enhances salient features. The second innovation is the Dynamic Feature Fusion Encoder (DFFE), a module adapted from dynamic convolutions, modified to incorporate complex values, thus effectively merging the inputs from high-frequency and low-frequency signal processes post individualized treatment. Lastly, the ComplexFractiReLU (CFReLU) activation function stands as a pioneering contribution, merging fractal theory with complex operations, dynamically adjusting the ReLU activation slope by estimating the fractal dimension of input data. This integration of advanced techniques and novel concepts aims to significantly enhance the capabilities in spectral synthesis for super-resolution imaging.

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

DHS2Net: Dynamic Hybrid Spectral Synthesis Network for Single Image Super Resolution

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Zihao He,
  • Haoxin Hu

摘要

In this paper, we introduce the Dynamic Hybrid Spectral Synthesis Network (DHS2Net), a novel dual-domain learning approach. The network innovatively separates high and low frequency components using convolutional layers with different strides, subsequently directing the low-frequency elements through a sequence of residual in residual blocks for spatial domain refinement, while employing complex-valued convolutions for high-frequency feature processing. An integral component of this architecture is the Feature Enhancement and Selection Unit (FESU), which dynamically selects and enhances salient features. The second innovation is the Dynamic Feature Fusion Encoder (DFFE), a module adapted from dynamic convolutions, modified to incorporate complex values, thus effectively merging the inputs from high-frequency and low-frequency signal processes post individualized treatment. Lastly, the ComplexFractiReLU (CFReLU) activation function stands as a pioneering contribution, merging fractal theory with complex operations, dynamically adjusting the ReLU activation slope by estimating the fractal dimension of input data. This integration of advanced techniques and novel concepts aims to significantly enhance the capabilities in spectral synthesis for super-resolution imaging.