基于峰值感知和多尺度约束的加权引导滤波器  

Weighted guided filter based on peak-aware and multi-scale constraints

作　　者：张全 刘海忠[1] ZHANG Quan;LIU Haizhong(College of Mathematics and Physics,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学数理学院,甘肃兰州730070

出　　处：《浙江大学学报（工学版）》2024年第10期2053-2061,共9页Journal of Zhejiang University：Engineering Science

基　　金：甘肃省优秀研究生“创新之星”项目(2023CXZX-592).

摘　　要：引导图像滤波无法保留锐利边缘,导致平滑图像出现结构过度模糊问题,为此提出新的加权引导图像滤波器.利用峰值感知加权提取图像的边缘和结构信息,通过多尺度约束提高所提滤波器的鲁棒性.基于图像方差信息将滤波损失函数的正则化项系数改进为自适应式.边缘感知平滑、图像细节增强、纹理去除平滑以及图像去噪领域的应用实验结果表明,所提滤波器在视觉效果、峰值信噪比以及结构相似度上均优于参与对比的引导图像滤波器.与边缘感知平滑实验中次优滤波器的峰值信噪比和结构相似度相比,所提滤波器的峰值信噪比平均高2.62 dB,结构相似度平均高0.0286.A new weighted guided image filter was proposed for the problem that guided image filtering fails to preserve sharp edges and leads to excessive structural blurring in smoothed images.Peak-aware weighting was utilized to extract edge and structural information from images,and the robustness of the proposed filter was improved by multi-scale constraints.The regularization term coefficients of the filter loss function were improved to an adaptive form based on the image variance information.Application experiments were carried out in edge-aware smoothing,image detail enhancement,texture removal smoothing,and image denoising,and the results showed that the proposed filter outperformed the guided image filters involved in the comparison in terms of visualization,peak signal-to-noise ratio,and structural similarity.Compared with the peak signal-to-noise ratio and structural similarity of the suboptimal filters of the edge-aware smoothing experiments,the peak signal-to-noise ratio was 2.62 dB higher on average,and the structural similarity was 0.0286 higher on average.

关 键 词：引导滤波器 图像平滑 峰值感知 多尺度约束 边缘保持 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]