In this paper, we propose a tensor-based non-local filtering technique for image and MRI denoising using Bayesian CP factorization (BCPF). This approach simply groups together similar sub-tensors (e.g., 3D tensors) selected from a noisy tensor and forms a 4D stack, then decomposes this stack into latent factors by employing BCPF, resulting in a filtered group of 3D sub-tensors. This procedure is repeated across the entire tensor in sliding window fashion to obtain the denoised result of original tensor. Our Bayesian CP factorization can learn CP-rank as well as noise variance solely from the observed noisy tensor data, which can also avoid overfitting problem by employing a fully Bayesian treatment for latent factor inference. The main advantage of our method is that the standard deviation of Gaussion noise can be automatically inferred and not necessary to be fixed. The experimental results on image and MRI denoising demonstrate the superiorities of our method in terms of flexibility and performance, as compared to other tensor-based denoising methods.
Jianting Cao hasn't uploaded this paper.
Let Jianting know you want this paper to be uploaded.
