Scholarlink

<p>Images captured under low-light conditions frequently exhibit poor visibility, low contrast, severe noise interference, and color distortion, significantly degrading the performance of downstream computer vision tasks. To mitigate these issues, a multi-module collaborative network for low-light image enhancement (LLIE) is proposed. A Low-Frequency-Assisted Noise Elimination Module (LNEM) suppresses noise during initial feature extraction, while a Cross-Scale Feature Interaction Module (CSFIM) enhances multi-scale information flow. Concurrently, a Semantic Context Mining Module (SCM) improves spatial–channel dependencies by integrating deep convolutional features with attention mechanisms. An asymmetric encoder–decoder architecture is constructed, into which the CSFIM and SCM are integrated to facilitate cross-scale interaction and semantic refinement. The encoder employs progressive dilated convolutions to capture local and global features, and the decoder adopts a progressive upsampling strategy for refined detail reconstruction. Training incorporates both spatial and frequency domain losses. Extensive evaluations on paired and unpaired datasets demonstrate the method’s efficacy, achieving 26.025&#xa0;dB PSNR and 0.939 SSIM on the LOL-V2-synthetic dataset—surpassing the second-best method by 0.7&#xa0;dB and 0.011, respectively. The approach also attains the optimal NIQE score of 3.768 on DICM and a competitive 3.104 on VV, confirming its capability for high-fidelity low-light image enhancement.</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

Low-Light Image Enhancement via Multi-Scale Feature Interaction and Attention Fusion

  • 期刊论文

  • 发表日期：

    2026年7月8日

查看全文
  • Guangyu Xu,
  • Yaozheng Wang

摘要

Images captured under low-light conditions frequently exhibit poor visibility, low contrast, severe noise interference, and color distortion, significantly degrading the performance of downstream computer vision tasks. To mitigate these issues, a multi-module collaborative network for low-light image enhancement (LLIE) is proposed. A Low-Frequency-Assisted Noise Elimination Module (LNEM) suppresses noise during initial feature extraction, while a Cross-Scale Feature Interaction Module (CSFIM) enhances multi-scale information flow. Concurrently, a Semantic Context Mining Module (SCM) improves spatial–channel dependencies by integrating deep convolutional features with attention mechanisms. An asymmetric encoder–decoder architecture is constructed, into which the CSFIM and SCM are integrated to facilitate cross-scale interaction and semantic refinement. The encoder employs progressive dilated convolutions to capture local and global features, and the decoder adopts a progressive upsampling strategy for refined detail reconstruction. Training incorporates both spatial and frequency domain losses. Extensive evaluations on paired and unpaired datasets demonstrate the method’s efficacy, achieving 26.025 dB PSNR and 0.939 SSIM on the LOL-V2-synthetic dataset—surpassing the second-best method by 0.7 dB and 0.011, respectively. The approach also attains the optimal NIQE score of 3.768 on DICM and a competitive 3.104 on VV, confirming its capability for high-fidelity low-light image enhancement.