摘要
针对经典的块匹配和三维滤波(BM3D)降噪算法中最为核心的噪声水平(方差)参数在使用中需要人工手动设置极大影响了降噪效果并限制了它的应用,提出了一种新的基于自然场景统计(NSS)的噪声水平特征矢量和支持向量回归(SVR)技术的快速噪声水平估计算法并应用于经典BM3D算法中,使之转变为自适应降噪算法(Adaptive BM3D)。本文算法首先利用小波变换对图像进行不同尺度和不同方向的分解,提取各子带滤波系数并用通用高斯分布模型(GGD)建模,以模型参数构成反映噪声图像噪声水平的特征矢量;然后用SVR方法在大量噪声图像样本上进行训练获得图像噪声水平预测模型。实验表明:改进后的ABM3D算法实际图像降噪效果比BM3D算法获得进一步提升,并且仍然保持了非常高的执行效率,相对于当前各主流算法具有明显的竞争力。
As the core parameter of the classical block-matching and 3D filtering （BM3D） algorithm, noise level （variance） needs to be manually set, which greatly affects the BM3D algorithm＇s noise reduction performance and limits its application. To resolve this problem, a fast noise level estimation algorithm that utilizes the feature vectors based on the natural scene statistics and support vector regression （SVR） techniques is proposed,based on which the standard BM3D algorithm is transformed into an a- daptive denoising algorithm （adaptive BM3D）. Specifically, the sub-band coefficients of an image obtained from a wavelet transform over three scales and three orientations are parameterized using a generalized Gaussian distribution （GGD） ,and these estimated parameters are used to form a feature vector for de- scribing image noise level of the image. Given a lot of feature vectors obtained from training noisy ima- ges, we utilize support vector regression （SVR） to train an estimation model to predict the noise level for any noisy image. Experimental results show that the actual image noise reduction capability of the pro- posed ABM3D algorithm is much better than that of the classical BM3D algorithm,and it still maintains high efficiency, which gives it a significant competitive edge compared with other existing state of the art algorithms.
出处
《光电子．激光》
EI
CAS
CSCD
北大核心
2017年第6期663-673,共11页
Journal of Optoelectronics·Laser
基金
国家自然科学基金(61662044
61163023
61379018)
国家级大学生双创项目(201510403030)资助项目
