CT Image Denoising with Non-Local Means Based on Feature Fusion

For the problem that computer tomography (CT) images after denoising by the non-local mean algorithm cause edge fog and the disappearance of small feature information, an adaptive non-local mean denoising method based on feature fusion is proposed. Firstly, the similarity judgment of the center pixel is carried out to exclude the effect of non-similar pixels on the denoising effect. Then a Gaussian weighting method based on feature fusion is proposed, considering the self-similarity of images from the maximum eigenvalue of similar frame matrix and Euclidean distance between pixels. Finally, the supremum and infimum of the adaptive filter coefficient are constrained based on the structure tensor, which solves the problem that image quality is affected when the infimum of filter coefficient is zero. Simulations and practical applications prove that the proposed algorithm has better edge protection and detail information effect. The proposed algorithm improves the structure similarity by about 4% on average, and the peak signal to noise ratio increases by nearly 4 dB on average, compared with the non-local mean algorithm.