| In the process of magnetic resonance (MR) imaging, it is inevitable tointroduce noise due to the imaging system. The noise of image will greatlyreduce the image quality, make the organization boundary fuzzy and finestructure difficult to identify, which will influence medical diagnosis. So inthe field of MR image denoising, it is of great importance to effectivelyremove addictive noise while retaining good boundary and structureinformation.The nonlocal (NL) means filter is a successful approach proposed inrecent years due to its patch similarities comparison. However, theaccuracy of similarities in this algorithm degrades when containing heavynoise. In this paper, based on the perception that folded gray and whitematters are commonly found in MR images, we introduce a multi-channelfilter based feature similarities into NL-means filter. The multi-bank basedfeature vectors of each pixel in the image are computed by convolvingLeung-Malik set (contains edge, bar, and spot filters) from variousorientations and scales, and then the similarities according to thisinformation are computed instead of pixel intensity.In this paper, we select the anisotropic diffusion denoising method,the wavelet based denoising method, and the traditional nonlocal means tocompare with the proposed new methods. In the experiments, we considerthree different types of noise, Gaussian noise, Rician noise and real noisewhich produced at different levels. Meanwhile, several indicators arechosen to measure different algorithms, including peak signal-to-noiseratio (PSNR), mean square error (MSE), structure similarity (SSIM), andlocal variance distribution of similarity (QILV). Through both subjective and objective evaluations of variousalgorithms, we draw the conclusion: the multi-channel filter bank basednonlocal means achieve better results when removing image noise, whichmeans better recovery of the edge and anatomic structure of the originalimage. The higher is the noise level, the more obvious is this advantage. |
PDF Full Text Request