基于几何纹理与Anscombe变换的蜂窝材料太赫兹图像降噪模型

为解决太赫兹（Terahertz，THz）图像内泊松高斯混合噪声导致芳纶纤维蜂窝材料脱粘缺陷轮廓检测精度低的问题，基于Anscombe变换与小波阈值法构建了THz图像降噪模型。高斯噪声方差为降噪模型的必要参数，但实际THz图像噪声分布未知，且噪声与纹理在高频混叠，给方差准确估计提出了挑战。为此，首先以样件纹理几何形状为先验信息，构造Benzene-ring算子去除THz图像纹理，使其小波域高频分量中仅含有噪声；然后提出改进的Logistic混沌映射提高样本集的多样性，以训练Elman神经网络准确建立高频分量与高斯噪声方差间映射关系；最后依据噪声方差估计值，基于Anscombe变换将泊松高斯混合噪声转化为高斯噪声，并利用小波阈值法与Anscombe逆变换得到了最终THz降噪图像。仿真与试验结果表明，所提出的方法降噪效果最佳并有效提高缺陷轮廓检测精度，相比于高斯滤波、小波阈值以及非局部均值法，平均梯度指标分别提升12%、33%、9%，缺陷面积绝对误差分别降低234 mm²、304 mm²、263 mm²。

Denoising Model of Terahertz Imaging of Honeycomb Material Based on Geometric Texture and Anscombe Transformation

In order to solve the problem of low accuracy caused by mixed Poisson-Gaussian noise in Terahertz (THz) images of aramid fiber honeycomb material in contour detection of defects, a THz image denoising model is constructed based on the Anscombe transformation and wavelet threshold method. The variance of Gaussian noise is a necessary parameter of denoising model, but the noise distribution of THz images is unknown. Meanwhile, the texture and noise are mixed in high frequency, which gives challenges to accurate variance estimation. Firstly, taking the texture geometry of the sample as the prior information, the Benzene-ring operator is constructed to remove the texture of the THz image, so that the high-frequency components only contain noise. Secondly, an improved Logistic chaotic mapping is proposed to improve the diversity of dataset to train Elman neural network for building the mapping relationship between the high-frequency component and the variance of Gaussian noise. Finally, the Anscombe transformation is performed to transform the mixed Poisson-Gaussian noise to gaussian noise. The THz denoised image is obtained by using wavelet threshold method and Anscombe inverse transformation. The experiment and simulation results show that the proposed model can eliminate the mixed Poisson-Gaussian noise better than other three methods and improves the accuracy of defect contour detections. Compared with the Gaussian filtering, wavelet threshold and non-local mean value methods, the average gradient index is increased by 12%, 33%, and 9%, and the absolute error of defect area is reduced by 234 mm², 304 mm², and 263 mm², respectively.