Journal of Computer Applications ›› 2015, Vol. 35 ›› Issue (9): 2661-2665.DOI: 10.11772/j.issn.1001-9081.2015.09.2661

Previous Articles     Next Articles

Weighted guided image filtering algorithm using Laplacian-of-Gaussian edge detector

LONG Peng, LU Huaxiang   

  1. Institute of Semiconductors, Chinese Academy of Sciences, Beijing 100083, China
  • Received:2015-04-01 Revised:2015-04-27 Online:2015-09-10 Published:2015-09-17

LoG边缘算子改进的加权引导滤波算法

龙鹏, 鲁华祥   

  1. 中国科学院 半导体研究所, 北京 100083
  • 通讯作者: 龙鹏(1990-),男,湖南邵阳人,硕士研究生,主要研究方向:彩色图像与医学图像增强、分割,longpeng@semi.ac.cn
  • 作者简介:鲁华祥(1965-),男,浙江绍兴人,研究员,博士,主要研究方向:半导体神经网络。
  • 基金资助:
    中国科学院战略性先导专项基金资助项目(XDA06020700)。

Abstract: The original guided image filter algorithm performs not robust enough because it occupies the same local linear model among all the local patches while ignoring the texture difference. Based on the absolute magnitude of LoG (Laplacian-of-Gaussian) strength, a locally adaptive weighting parameter was used to penalize the fixed regularization parameter to produce a more robust method, aiming to amplify the grey scale difference between flat patch and edge patch, meanwhile avoid degraded denoising performance of original method. The open medical database BrainWeb including 6 T1, 6 T2 and 6 PD weighted pictures added with 9% magnitude of Racian noise were used as the testing database. Structural Similarity Index Measurement (SSIM) and Cumulative Probability of Blur Detection (CPBD) were used as quantity value indexes. According to the best experiment results, the proposed method respectively gets 5% and 6% advancement for SSIM and CPBD, compared to original guided image filter algorithm. Furthermore, the proposed method performs better than both the original guided image filter and another improved guided image filter under each regularization parameter of guided image filter, and the original O(N) time complexity is not affected. Compared to state-of-the-art methods, the proposed method obtains best performance compromising SSIM and CPBD, and it has lowest time complexity, while providing a fast and robust denoising method for medical images and color images.

Key words: guided image, edge-preserving filter, Laplacian-of-Gaussian (LoG) edge detector, local linear model, parameter self-adaption

摘要: 针对原始全局的引导滤波算法对整幅图像各个区域使用统一的线性模型与相同的规整化因子,从而未能适应图像本身不同区域的纹理特性,提出了基于LoG边缘检测算子改进的加权自适应规整因子。通过在局部窗口内计算LoG幅值响应,对原有的规整化因子进行惩罚来取得对图像平滑区域与边缘区域的自适应,使得在保证降噪效果的前提下进一步突出边缘像素和平坦区域像素之间的差异。对开源医学图像库BrainWeb中不同断层的T1、T2与PD加权图像,共18张图像,添加9%的莱斯噪声作为测试库,并采用结构相似性因子(SSIM)与无参考图像锐化因子(CPBD)作为算法的定量评估指标。实验结果表明,与原始的引导滤波算法相比,所提方法的SSIM指标获得了最高5%左右的提升,CPBD指标获得了最高6%左右的提升。在引导滤波不同规整化因子的条件下,所提算法均优于原始的引导滤波算法和现有的基于方差图像加权改进的引导滤波算法,并保留了原始引导滤波O(N)的复杂度。与现存的主流滤波算法比较,所提算法能够兼顾SSIM与CPBD指标,具有最高的综合性能,且具有最低的算法复杂度,能够用于医学图像和彩色图像的快速滤波降噪。

关键词: 引导图像, 边缘保持滤波, LoG边缘检测算子, 局部线性模型, 参数自适应

CLC Number: