变分模态分解与稀疏SURE的电子图像噪声抑制

Noise suppression for electronic images using variational mode decomposition and sparse SURE algorithm

摘要:

为解决电子微结构图像在摄取、传输或存储的过程中易被外界噪声干扰、图像保真度差的问题,提出了一种变分模态分解与稀疏Stein无偏风险估计(Stein unbiased risk estimator,SURE)相结合的图像噪声抑制方法,以铝合金、双相钢与钛合金Ti6Al4V 3种材料的电子背散射衍射图像为例. 首先,在已采集的电子背散射衍射图像中加入高斯噪声与Speckle斑纹噪声来模拟被干扰图像；然后,利用变分模态分解方法按照频率尺度将含噪模拟图像分解为固有特征成分与高频噪声成分；继而利用Haar小波冗余字典对固有特征成分进行稀疏表示,在一阶可导收缩函数的基础上推导了稀疏Stein无偏风险估计阈值选择的优化目标函数,最后,利用黄金分割搜索法计算得到全局最佳自适应阈值. 结果表明:提出的方法可有效去除外界干扰噪声,提高了图像的峰值信噪比；以铝合金为例,当噪声标准差为30时,提出方法的图像峰值信噪比突破了单一稀疏SURE收缩曲线的最大值,比Neigh-Shrink方法高0.39 dB,比KSVD方法高2.895 dB,比小波阈值去噪算法高3.07 dB.

Abstract:

Electronic images are easily contaminated and blurred during the processes of acquisition, transmission, and storage, and the fidelity of the original images are degraded. To address this issue, a hybrid denoising method based upon variational mode decomposition (VMD) and Stein unbiased risk estimator (SURE) was proposed, taking the electron backscatter diffraction (EBSD) images of Aluminum alloy, dual-phase steel, and Ti6Al4V as examples. To begin with, the clean EBSD images were contaminated by adding Gaussian noise and speckle patterns noise. Then, the noisy image was decomposed into characteristic information component and high-frequency noise components using the Bi-variational mode decomposition (BVMD) algorithm. Next, the generated inherent characteristic component was fed into the Haar wavelet redundant dictionary (HWRD) for sparse representation. Meanwhile, the optimal objective shrinkage function was derived with the function of one-order differentiable shrinkage. Finally, the adaptive threshold was obtained via golden section search (GSS) method. Experimental results show that the proposed method could effectively remove the external interference noise and improve the peak signal-to-noise ratio (PSNR) of the image. Specifically, taking the EBSD image of Aluminum alloy as an example, when the noise standard deviation was 30, the proposed method exceeded the maximum point of single sparse SURE method in terms of PSNR value, which also outperformed the Neigh-Shrink method by 0.39 dB, the K-singular value decomposition (KSVD) method by 2.895 dB, and the soft-wavelet threshold denoising (SWTD) by 3.07 dB.