Research On Denoising Methods For Laser Speckle Imaging

Laser speckle contrast imaging(LSCI)is a fast,wide-field,and noncontact imaging technology for mapping blood flow.However,contrast calculation is affected by various noises especially in real-time imaging.The existing denoising methods of LSCI mainly adopt spatiotemporal average and local filtering,which lack of using non-local information in the image.The denoised images still contain noise.Real-time and high-quality denoising is also not achieved.In addition,laser speckle images captured by LSCI can also be used for multi-modal imaging such as tissue structure and blood oxygen saturation.Speckle reduction is needed to improve imaging accuracy.When only laser illumination is used,the existing self-supervised learning method,Noise2Noise(N2N),can not obtain two speckle images with independent noise for training.As a result,N2 N remains speckle noise.This thesis focused on three problems of existing denoising methods: lack of using non-local similarity information of LSCI,lack of high-quality and real-time denoising method for LSCI,and laser speckle noise remained by N2 N.The main contents were listed as following:(1)To improve the denoising quality by using the non-local similarity information of LSCI images,Manhattan distance-based adaptive block-matching and three-dimensional collaborative filtering(MD-ABM3D)is proposed.The local noise variance is estimated by the adaptive algorithm,so that the filtering can adapt to the noise in different velocity regions,which improves 2.09 d B in peak signal-to-noise ratio(PSNR).Manhattan distance is adopted to improve the stability of block-matching against strong noise in inhomogeneous noise,which further increases 0.09 d B.The image-quality evaluations of MD-ABM3 D for temporal LSCI(t LSCI)(temporal window Wt = 20 frames)is better than that of spatially averaged t LSCI(Wt = 60 frames).(2)To further realize high-quality and real-time denoising,a dilated deep residual learning network with skip connections(DRSNet)in the log-transformed domain is proposed.The reference image is obtained by denoising t LSCI(Wt = 50 frames)through MD-ABM3 D,which is used for network learning to denoise spatiotemporal LSCI(Wt =1～10 frames).Denoising convolutional neural network(Dn CNN)can effectively remove the conventional additive Gaussian noise,but it remains noise when processing squared contrast images with multiplicative noise.By converting the images to the log-transformed domain,the distribution of noise level in images is more uniform,which improves 5.13 d B in PSNR.After reducing the numbers of layers and convolution channels of Dn CNN,the receptive field is increased through dilated convolution,and the multi-scale information is fused by skip connection.DRSNet further improves 0.15 d B while its denoising speed is2.5 times faster than Dn CNN.DRSNet takes 35 ms(28 fps)for denoising a blood flow image with 486 × 648 pixels on an NVIDIA 1070 GPU,which realizes real-time denoising and high denoising quality.(3)To remove the speckle noise remained by N2 N,anisotropic gradient regularizationbased Noise2Noise(AGR-N2N)in the log-transformed domain is proposed for laser speckle denoising in living organisms.By converting laser images to the log-transformed domain,the denoising robustness of N2 N is enhanced.It improves 0.59 d B in PSNR.Anisotropic gradient regularization(AGR)accurately removes the speckle noise remained by N2 N,while retains the connected blood vessels and other tissue structures,which further improves 2.10 d B.The AGR calculation is speed up by convolutions.The edge effect of AGR is removed by padding infinity at the edges.AGR-N2 N takes 21 ms(48 fps)to denoise an image with 512 × 512 pixels on an NVIDIA RTX 3090 GPU.