图像去噪专题讲座.ppt

Natural images have a great deal of redundancy: patches from different regions can be very similar. Non-Local Means Buades et al., 2005 (CVPR) （被引用次数：2252） Is denoising dead? 56/66 Non-local means: Basic Principle Non-local means compares entire patches (not individual pixel intensity values) to compute weights for denoising pixel intensities. Comparison of entire patches is more robust, i.e. if two patches are similar in a noisy image, they will be similar in the underlying clean image with very high probability. 57/66 Non-Local Means Buades et al., 2005 (CVPR) How-to: NL-Means It turns out this can be formulated in the Gaussian filtering framework. 58/66 From Bilateral Filter to NLM intensity weight spatial weight Patch similarity 59/66 clear clc ima=double(imread(house2.jpg)); % add some noise sigma=10; rima=ima+sigma*randn(size(ima)); % denoise it fima(:,:,1)=NLmeansfilter(ima(:,:,1),5,2,sigma); fima(:,:,2)=NLmeansfilter(ima(:,:,2),5,2,sigma); fima(:,:,3)=NLmeansfilter(ima(:,:,3),5,2,sigma); % show results subplot(1,3,1),imshow(ima/255,[]),title(original); subplot(1,3,2),imshow(rima/255,[]),title(noisy); subplot(1,3,3),imshow(fima/255,[]),title(filtered); 60/66 Ref: http://demo.ipol.im/demo/bcm non local means denoising/ Codes: /matlabcentral/fileexchange/13176-non-local-means-filter 61/66 /question跟其他被Deep Learning刷PSNR到新高的low-level vision领域，比如Super-resolution, Inpainting, Deblocking这些相比，在Denoising这个方向上，Deep Learning罕见地没有大幅度地刷高baseline。  62/66 Beyond ? 作业4：双边滤波实现及应用 Bilateral Filtering: Theory and Applications A Gentle Introduction to Bilateral Filtering and its Applications /sparis/bf_course/ 63/66 作业5: nonlocal均值方法实现及其应用 https://www.iro.umontreal.ca/~mignotte/IFT6150/Articles/Buades-NonLocal.pdf 3. Is denoising dead? 4. A review of image denoising algorithms, with a new one 64/66 图像去噪是典型的病态反问题，如果一个新模型想在图像复原领域里得到认可，那么这个模型首先要在图像去噪这个问题上试一把。此外，求解中通常将图像复原模型分解为一些简单的子问题逐个求解(各个击破)，图像去噪是其中最重要的子问题。 65/66 总结 Plug and Play ADMM for Image Restorati