Abstract: This paper proposes an algorithm improved by anisotropic diffusion filtering. Existing research methods have many problems, such as unclear with the image edge, also there is a plenty of false recognition and it is usually a subjective choice to the diffusion coefficient. The algorithm uses Median Filtering Method in multi directions of diffusion which can retain the edge and the details. Improve the diffusion coefficient by using local variance and image gradient and modify the diffusion coefficient by multiple iterations, and this can enhance the robustness of the algorithm. The algorithm can de-noise and retain the edges and the details. By several simulation experiments, compared the simulation results with the peak signal noise ratio, the mean square error, the structure similarity and the figure of merit as four quantitative indicators. It shows that the algorithm has great ability on de-noise and retain image edges than the traditional anisotropic diffusion method and Catt _PM model.

Key words: anisotropic diffusion filtering, median filtering method in multi directions, diffusion coefficient, local variance, image gradient

摘要： 提出一种改进各向异性扩散滤波算法。现有研究方法多存在图像边缘不清，误识别多，扩散系数多凭主观选择等问题。该算法利用保留细节和边缘的能力较为突出的多方向中值滤波方法在多个方向上进行扩散，利用局部方差和图像梯度改进了扩散系数，通过多次迭代修正扩散系数，增强了算法的鲁棒性，且在滤除噪声的同时注重对边缘细节的保持。通过具体实验仿真，以峰值信噪比、均方误差、结构相似度以及图像佳数4个参数作为指标，对实验仿真结果进行了量化比较，表明该算法与传统各向异性扩散方法以及Catté_PM模型等改进方法相比，具备更好的滤除图像噪声以及保持图像边缘的能力。

关键词: 各向异性扩散滤波, 多方向中值滤波, 扩散系数, 局部方差, 图像梯度