基于HVL色彩空间与傅里叶核估计实现低光模糊图像恢复

Low-light Blurry Image Restoration via HVL Color Space and Fourier Kernel Estimation

摘要:

针对低光照与运动模糊交织的复合降质修复问题，提出了一种低光模糊图像修复网络，旨在提升复杂环境下图像质量。该方法通过引入水平/垂直/光强颜色空间，利用极化处理色调与饱和度平面及自适应亮度坍缩函数，解决低光增强中的色彩断裂与黑色平面噪声问题。构建了频域核估计模块，通过傅里叶空间中的卷积过程精确建模物理模糊特征。同时，采用全局Mamba块作为核心单元，提升了长程依赖建模能力并降低了计算负担。经LOLBlur和Real-LOLBlur数据集测试，结果表明该方法在峰值信噪比上提升约1dB，图像块相似性指标降低近50%，并有效控制了参数量与计算开销，实现了性能与推理效率的平衡

Abstract:

A low-light motion blur image restoration network is proposed to address the composite degradation caused by low light and motion blur, aiming to enhance image quality in complex environments. The method introduces horizontal/vertical/intensity color spaces, utilizing polarization processing for hue and saturation planes and an adaptive brightness collapse function to address color break and black-plane noise in low-light enhancement. A frequency-domain kernel estimation module is constructed to precisely model physical blur characteristics through convolution in the Fourier space. Additionally, a global Mamba block is employed as the core unit to improve long-range dependency modeling while reducing computational burden. Experimental results on the LOLBlur and Real-LOLBlur datasets demonstrate that the method improves peak signal-to-noise ratio by approximately 1dB, reduces the image block similarity index by nearly 50%, and effectively controls parameter count and computational cost, achieving a balance between performance and inference efficiency.