提出了一种基于张量的平稳小波变换红外图像去噪方法。采用平稳小波对噪声红外图像进行分解, 保持低频近似图像不变, 将所有尺度上的水平、垂直和对角方向的高频细节图像组合为一个立方体, 形成三阶张量, 通过多线性代数方法估计信号小波系数, 这种处理方式没有破坏小波系数之间的固有空间关系, 同时考虑到了尺度间和尺度内小波系数的相关性, 优于传统的基于线性最小均方误差的信号小波系数估计算法, 最后由低频近似图像与估计的高频细节图像通过平稳小波逆变换得到去噪图像。实验结果表明, 该方法在性能指标和视觉质量上优于传统的平稳小波域最小均方误差去噪算法, 为小波系数的较准确估计提供了一种全新思路。

A method based on stationary wavelet transform using tensor for infrared image denoising was proposed. The noisy infrared image was decomposed by stationary wavelet transform. The low frequency approximation image was not changed. A cube was constituted by the high frequency sub-band images at horizontal, vertical and diagonal directions of the whole scales. And a third-order tensor was formed. The signal wavelet coefficients were estimated by multi-linear algebra method. The space structural relations of wavelet coefficients were not destroyed in this way. The correlations of coefficients both inter-scale and intra-scale were considered at the same time. The wavelet coefficients which were estimated in this method were better than the linear minimum mean square-error estimation (LMMSE) method. The denoised image was obtained by inverse stationary wavelet transform using the low frequency approximation image and the estimated high frequency detail images. Experiment result shows that the denoising results were better than the traditional LMMSE estimation with stationary wavelet domains in performance evaluation and visual quality. And a new thought was provided to estimate the wavelet coefficients more accurately.