Research On Despeckling Of OCT Image Based On The Regularization

With the development of medical imaging,the imaging technologies have been popping up quickly.The department of ophthalmology benefits from the emergy and development of the optical coherence tomography(OCT)imaging technology very much,and it has been widely used.But inevitably the speckle noise arises because the mode of coherent imaging,and it seriously affects the subsequent processing of the OCT image and the diagnosis of disease correctly.Image denoising as preprocessing method,it has been applied in medical images,and has its own uniqueness and challenging at the same time.Based on the theory of regularization,the despeckling algorithms of OCT image are studied futher in this thesis which the main contents are as follows:(1)A despeckling algorithm for OCT image based on total variational regularization is proposed.The first,we get the OCT image which speckle noise is subject to the gamma distribution according to analysis and prove the distribution of speckle noise through the histogram.Then according to the distribution of speckle noise,the data fidelity term is established.Combining the total variational regularization term,the objective function is builded.The alternating direction method of multipliers(ADMM)method is applied to solve the constrained optimization problem caused by objective function quickly after variable splitting.A series of experiments are performed on the proposed method for real OCT images,and the algorithm is evaluated by image quality evaluation indexes,which show that the proposed algorithm not only reduces the noise and retains the detail and edge information very well,but also it runs very quickly,the objective function can convergence in a relatively short time.(2)Hyper-laplacian priors is applied to the despeckling algorithm.The part is on the basis of total variational regularization model which is proposed in previous part,retains the data fidelity term.For the regularization term,we found that the gradient image of OCT have heavy tailed phenomena,so p norm is adopted as the regularization term to fit the prior knowledge of the image in this part.Then combining the fidelity term establishes the objective function instead of total variation model.But the p norm leads to the convex optimization problem into the non convex problem which is more difficult to solve.by analysis and summary,this thesis uses quadratic penalty method and look-up table to solve non convex problem.At last,compared with the proposed algorithm in the previous,the result of the experiments shows that the proposed algorithm obtains the better result by using the correct prior knowledge.