A geometry guided image denoising scheme

During image denoising, it is often diffcult to balance between the removal of noise and the preservation of contrast and fine features, especially when the noise is excessive. We propose to efficiently balance the two using segmentation and more general geometry extraction transforms. Explained in the nonlocal-means (NL-means) framework, we introduce a mutual position function to ensure the averaging is only taken over pixels in the same segmentation phase, and provide selection schemes for convolution kernel and weight function to further improve the performance. To address unreliable segmentation due to more excessive noise, we use a feature extraction transform that is more general than segmentation and less sensitive to noise. Unlike most de- noising approaches that only work for one type of noise and/or involve heuristic parameter tuning, the proposed method comes with an automatic parameter selection scheme, and can be easily adapted for various types of noise, ranging from Gaussian, Poisson, Rician to ultrasound noise. Comparison with the original NL-means as well as ROF, BM3D, and K-SVD on various simulated data, MRI and SEM images, indicates potentials of the proposed method.