| 引用本文: | 王科俊,熊新炎,任桢,付斌.基于窗口无关均值滤波的MSR图像增强[J].哈尔滨工业大学学报,2010,42(9):1467.DOI:10.11918/j.issn.0367-6234.2010.09.024 |
| WANG Ke-jun,XIONG Xin-yan,REN Zhen,FU Bin.MSR image enhancement based on window irrelevant mean filter[J].Journal of Harbin Institute of Technology,2010,42(9):1467.DOI:10.11918/j.issn.0367-6234.2010.09.024 |
|
| 摘要: |
| 为解决多尺度Retinex方法在运算中要多次运用高斯卷积,尤其是采用大尺度高斯核进行卷积获取图像低频信息时,运算速度将大大下降这一问题,根据大尺度高斯卷积模板具有均值模板的特性,利用量化的均值滤波近似大尺度高斯卷积,将卷积模板简化成均值模板,并提出一种与滤波窗口无关的快速均值算法.该方法引入了一个数组s用于存储窗口横向移动时每一列的数值,从而减少重复计算,大大提高滤波速度.以640×640大小的彩色图像为例,其处理速度仅为0.0867s,相比原始的MSR算法速度提高了732倍.试验分析表明,该方法即保证实现良好的图像增强效果,同时极大的提高了运算效率. |
| 关键词: 多尺度Retinex 图像增强 窗口无关均值滤波 高斯卷积 |
| DOI：10.11918/j.issn.0367-6234.2010.09.024 |
| 分类号:TP391.41 |
| 基金项目:国家高技术研究发展计划资助项目(2008AA01Z148) |
|
| MSR image enhancement based on window irrelevant mean filter |
|
WANG Ke-jun, XIONG Xin-yan, REN Zhen, FU Bin
|
|
College of Automation,Harbin Engineering University,Harbin 150001,China
|
| Abstract: |
| To increase the computation efficiency of multi-scale retinex method(MSR) in image enhancement,in which the Gaussian convolution has to be used for several times,especially when large scale Gaussian convolution template is used to get the low-frequency information of image,the convolution template is simplified as mean template because the Gaussian template of large scale Gaussian convolution has the property of mean template.Furthermore,a filter window irrelevant mean filtering algorithm is proposed to fulfill and speed up the filtering process.It employs an array s to storage each column pixel value of current filtering rows and consequently reduces the redundant calculation.Taking the 640 × 640 color image as an example,the processing time is only 0.0867 s and it is 723 times faster than original MSR method.It can be seen from the analysis of experiment that the new proposed method can greatly improve the computing efficiency and fulfill the image enhancement remarkably. |
| Key words: multi-scale retinex image enhancement window irrelevant mean filter Gaussian convolution |
