Application Of Non-local Means In Low Signal-to-Noise Ratio CT Imaging

Computed tomography(CT)technology uses the attenuation information when the X-ray passes through the object to be measured to get the density distribution of the object,and it can clearly present the internal structure information of the object.CT non-destructive testing requires the reconstruction image should have high signal-to-noise ratio(SNR)and resolution.However,in the practical application of CT technology,due to the different detection requirements,different application scenarios or different CT systems,there will be a lot of noise and artifacts in the reconstructed image,and the low SNR of the image will seriously affect the subsequent image processing,such as image segmentation and high-precision measurement.Therefore,it is of great academic significance and practical value to study the filtering method to improve the quality of CT image with low SNR.Aiming at the problems of denoising the low SNR image in industrial CT imaging,we focused on the filtering effect of the Non-local Means(NLM)algorithm on the projection data and the CT reconstructed images,and proposed improved algorithms.The main research contents are as follows:Firstly,studied and implemented three classical noise reduction algorithms:anisotropic diffusion filtering,bilateral filtering and NLM,after then analyzed their advantages and disadvantages.Simulation experiments and actual CT scan data show the superiority of NLM algorithm in improving the quality of CT images with low SNR,which can preserve the edge information.Secondly,a NLM algorithm based on singular value decomposition(SVD-NLM)was proposed.For the misjudgment of similarity is often caused by noise interference in the calculation of image block similarity by NLM algorithm which was applied in low SNR CT image filtering.So as to find the similar block more accurately,singular value decomposition was added into the evaluation of image block similarity to reduce the noise interference effectively.At the same time,aiming at the time-consuming problem of NLM algorithm,the integral image was used to accelerate the NLM algorithm.The feasibility and effectiveness of the algorithm were verified by experiments,that was,SVD-NLM algorithm can retain more edge details and other useful information and get better noise reduction performance,and the running speed of the algorithm was about twice faster than that of NLM algorithm.Thirdly,a NLM algorithm based on structure tensor(ST-NLM)was proposed.Filtering coefficient is an important factor that affects the performance of NLM algorithm.Aiming at the problem that the details of CT images with low SNR are too smooth or not smooth enough due to the use of fixed filtering coefficient in NLM algorithm,an adaptive filtering coefficient was proposed.The structure tensor is introduced to represent the geometric structure information of the image.Qualitative and quantitative experiments demonstrated that ST-NLM algorithm can effectively reduce the noise while maintaining the edge information.Compared with NLM algorithm,the peak signal-to-noise ratio,structure similarity and the running speed of ST-NLM algorithm increase by 5%,3d B and twice,respectively.