Scholarlink

<p>Low-light image enhancement is a key task in low-level computer vision, aiming to improve visibility while preserving structural and textural details. Most existing Transformer-based methods achieve notable progress by modeling global dependencies but still struggle to recover fine-grained high-frequency details due to limited local perception. To address this, we propose the Wavelet-Guided Dual-Branch Cross-Attention Fusion Network (WDCFNet), which employs a horizontal–vertical dual-path scanning strategy inspired by state space models. In this design, the dual-path masks are used as explicit positional encodings and embedded into the Vision Transformer, enhancing both local detail capture and long-range semantic modeling. To further preserve structure and texture, the low-light input image is first converted into the Horizontal/Vertical-Intensity (HVI) color space and split into HV and I branches. In parallel, the wavelet-guided dual-branch reconstruction module decomposes the input into LL, HL, LH, and HH subbands. The LL subband guides the I branch for structure and illumination restoration, while the fused high-frequency subbands enhance the HV branch for texture recovery. The outputs are finally fused via channel-wise integration to generate the enhanced image. Extensive experiments demonstrate that WDCFNet significantly outperforms existing state-of-the-art methods in both quantitative and qualitative evaluations. The code is available at <a href="https://github.com/mlxgccc/WDCFNet">https://github.com/mlxgccc/WDCFNet</a>.</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

WDCFNet: a wavelet-guided dual-branch cross-attention fusion network for low-light image enhancement

  • 期刊论文

  • 发表日期：

    2026年4月11日

查看全文
  • Shuqin Zhang,
  • Kaiyu Li,
  • Xueqing Wang,
  • Yuanqing Xia

摘要

Low-light image enhancement is a key task in low-level computer vision, aiming to improve visibility while preserving structural and textural details. Most existing Transformer-based methods achieve notable progress by modeling global dependencies but still struggle to recover fine-grained high-frequency details due to limited local perception. To address this, we propose the Wavelet-Guided Dual-Branch Cross-Attention Fusion Network (WDCFNet), which employs a horizontal–vertical dual-path scanning strategy inspired by state space models. In this design, the dual-path masks are used as explicit positional encodings and embedded into the Vision Transformer, enhancing both local detail capture and long-range semantic modeling. To further preserve structure and texture, the low-light input image is first converted into the Horizontal/Vertical-Intensity (HVI) color space and split into HV and I branches. In parallel, the wavelet-guided dual-branch reconstruction module decomposes the input into LL, HL, LH, and HH subbands. The LL subband guides the I branch for structure and illumination restoration, while the fused high-frequency subbands enhance the HV branch for texture recovery. The outputs are finally fused via channel-wise integration to generate the enhanced image. Extensive experiments demonstrate that WDCFNet significantly outperforms existing state-of-the-art methods in both quantitative and qualitative evaluations. The code is available at https://github.com/mlxgccc/WDCFNet.