基于信息几何的图像去噪

提出一种基于信息几何的图像去噪方法,与传统的欧氏空间中去噪方法不同,基于信息几何的图像去噪方法是在流形上利用两点间的测地线距离的大小来衡量图像中两像素点之间的相似性。流形中的点是通过图像中某区域构建的高斯模型,模型间的测地线距离表示两图像区域平均灰度强度和细节丰富程度的差别。这样将区域的平均灰度强度和细节丰富程度的差别作为像素点间的相似性,能够更加准确地衡量像素点间相似性的差异。实验表明,该算法提高了图像去噪效果,能够很好地保持图像细节。

Image denoising based on information geometry

We present an image denoising method based on information geometry. Different than the traditional Euclidean space image denoising, the image denoising method based on information geometry makes use of manifold geodesic distance between two points to measure the similarities between two pixels in an image. The points on manifold are built by the Gaussian model of a region in an image and the geodesic distance between models represents the differences of average gray intensity and the detailed richness of the two image regions, in which way the similarities between pixels could be measured more accurately. Experiments show that the algorithm improves the denoising effect and maintains good image details.