Scholarlink

<p>This paper proposes an Illumination-Adaptive Feature Modulation and Frequency-Spatial Cross-Domain Fusion Network (IAFM-FSNet) to address the limitations of static convolution kernels in adapting to spatial illumination variations, single spatial-domain processing overlooking frequency information, and simple feature fusion creating information conflicts in low-light image enhancement. Our network creatively creates an Illumination Gate module that uses a lightweight perception network to produce illumination-aware scaling factors in order to dynamically modify convolution responses at the lowest possible parameter cost. We build parallel frequency and spatial branches: the frequency branch uses learnable frequency filters to segregate and process high- and low-frequency information, while the spatial branch uses dual attention mechanisms and Illuminance-Adaptive Feature Modulation to capture spatial structural elements. A bidirectional cross-attention fusion module is suggested to create dynamic mutual guiding between the spatial and frequency domains. We construct parallel frequency and spatial branches: the frequency branch separates and processes high- and low-frequency information via learnable frequency filters, while the spatial branch captures spatial structural features through multi-scale Illuminance-Adaptive Feature Modulation and dual attention mechanisms. A bidirectional cross-attention fusion module is proposed to establish dynamic mutual guidance between frequency and spatial domain features, achieving deep information interaction and complementary enhancement. Finally, a multi-scale progressive enhancement strategy is employed to progressively optimize feature representations. Experimental results on three mainstream datasets (LOLv1, LOLv2_real, and LOLv2_synthetic) demonstrate that IAFM-FSNet achieves PSNR of 25.90&#xa0;dB and SSIM of 0.940 with only 0.54&#xa0;M parameters, significantly outperforming state-of-the-art methods such as Zero-DCE and MIRNet. Our method effectively solves problems including uneven brightness, detail loss, and color distortion, achieving an excellent balance between brightness improvement, detail preservation, and color naturalness, making it suitable for practical deployment in resource-constrained scenarios.</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

IAFM-FSNet: Illumination-Adaptive Feature Modulation and Frequency-Spatial Fusion for Low-Light Enhancement

  • 期刊论文

  • 发表日期：

    2026年5月22日

查看全文
  • Dong Guo,
  • Yi Han,
  • Yang Cui

摘要

This paper proposes an Illumination-Adaptive Feature Modulation and Frequency-Spatial Cross-Domain Fusion Network (IAFM-FSNet) to address the limitations of static convolution kernels in adapting to spatial illumination variations, single spatial-domain processing overlooking frequency information, and simple feature fusion creating information conflicts in low-light image enhancement. Our network creatively creates an Illumination Gate module that uses a lightweight perception network to produce illumination-aware scaling factors in order to dynamically modify convolution responses at the lowest possible parameter cost. We build parallel frequency and spatial branches: the frequency branch uses learnable frequency filters to segregate and process high- and low-frequency information, while the spatial branch uses dual attention mechanisms and Illuminance-Adaptive Feature Modulation to capture spatial structural elements. A bidirectional cross-attention fusion module is suggested to create dynamic mutual guiding between the spatial and frequency domains. We construct parallel frequency and spatial branches: the frequency branch separates and processes high- and low-frequency information via learnable frequency filters, while the spatial branch captures spatial structural features through multi-scale Illuminance-Adaptive Feature Modulation and dual attention mechanisms. A bidirectional cross-attention fusion module is proposed to establish dynamic mutual guidance between frequency and spatial domain features, achieving deep information interaction and complementary enhancement. Finally, a multi-scale progressive enhancement strategy is employed to progressively optimize feature representations. Experimental results on three mainstream datasets (LOLv1, LOLv2_real, and LOLv2_synthetic) demonstrate that IAFM-FSNet achieves PSNR of 25.90 dB and SSIM of 0.940 with only 0.54 M parameters, significantly outperforming state-of-the-art methods such as Zero-DCE and MIRNet. Our method effectively solves problems including uneven brightness, detail loss, and color distortion, achieving an excellent balance between brightness improvement, detail preservation, and color naturalness, making it suitable for practical deployment in resource-constrained scenarios.