Research On Noise Intensity Estimation And Its Application In Neutron Image Denoising

Neutron imaging technology is an important nondestructive testing technology,which utilizes images to display the internal structure of the object visually.It is widely used in many fields,such as national defense and medicine.However,in the process of imaging,transmission and storage,the neutron image formed by the small neutron imaging system with low fluence rate is inevitably disturbed by various noises,which result in severe degradation and seriously affect the subsequent image processing and detection efficiency.The most common model for noise is the additive white Gaussian noise.Currently,many high-performance denoising methods,such as three-dimensional block matching filtering(BM3D)and Non-local Means Filter(NLM),usually assume that the noise intensity is known,and the denoising performance is largely dependent on the noise intensity.But in fact the noise intensity is unknown,and it needs to be estimated.Therefore,it is very important to research reasonable and effective noise intensity estimation methods and apply them to denoising methods to achieve good denoising effect.Noise intensity is not easily to be estimated in degraded neutron images with low contrast.The existing noise estimation methods are applicable to some specific images,but tend to underestimate the noise intensity for the neutron image.Moreover,with the estimated noise intensity,denoising methods can not achieve the best performance.In view of this situation,this paper proposes two different noise intensity estimation methods and combines the classical BM3 D to perform denoising.In order to solve the low estimation problem and better preserve the details of image texture,this paper proposes a new noise intensity estimation method based on principal component analysis(PCA)and BM3 D,which is PCA-BM3 D for short.In this method,the residual image is introduced as the noise image.The maximum eigenvalue of the residual image obtained by PCA is used for pre-estimating noise intensity,which avoids the problem of low estimation.BM3 D utilizes the pre-estimated value for pre-denoising to preserve image details.Through the iterative process,the image information in the residual image is gradually reduced,and the eigenvalue of the residual image is minimized.According to the set stopping criterion,an accurate noise estimation value and a better denoising effect are finally obtained.To improve the processing speed and solve the problem that the existing estimation methods can not effectively select the weak texture regions of neutron images,another noise intensity estimation method based on structure tensor and residual image is proposed,which is ST-R for short.Firstly,based on the uniformity of image structure and structure tensor,the pre-selection,precise selection and labeling of weak texture regions are quickly completed,which reduces the computational complexity.Then,all the eigenvalues of the residual image in the corresponding region are used to estimate the noise intensity,so as to avoid the influence of the image information on the noise estimation and ensure the estimation accuracy.Combining BM3 D,it can get a good denoising effect.The experimental results show that compared with the existing methods,the two proposed methods can estimate the noise intensity more quickly and effectively.And the two methods combine BM3 D,both have satisfactory denoising effect,no matter visual perception or PSNR values.They have universal applicability and good robustness.They are of great significance to improve the practicability and reliability of neutron imaging technology and provide feasible support for the development of miniaturized neutron imaging equipment.