Filter bank based nonlocal means for denoising magnetic resonance images

Image denoising is a classical problem in image processing. Its essential goal is to preserve the image features and to reduce noise effectively. The nonlocal means (NL-means) filter is a successful approach proposed in recent years due to its patch similarity comparison. However, the accuracy of similarities in this algorithm degrades when it suffers from heavy noise. In this paper, we introduce feature similarities based on a multichannel filter into NL-means filter. The multi-bank based feature vectors of each pixel in the image are computed by convolving from various orientations and scales to Leung-Malik set (edge, bar and spot filters), and then the similarities based on this information are computed instead of pixel intensity. Experiments are carried out with Rician noise. The results demonstrate the superior performance of the proposed method. The wavelet-based method and traditional NL-means in term of both mean square error (MSE) and perceptual quality are compared with the proposed method, and structural similarity (SSIM) and quality index based on local variance (QILV) are given.