Fast nonlocal remote sensing image denoising using cosine integral images

A fast nonlocal means (NLM) filtering scheme, which uses cosine integral image, is proposed to reduce the computation cost of the standard NLM method. In the proposed method, the image patch similarity is estimated within the mean values of image patches, which are calculated by the summed image (SI) method. The weight function of the NLM is decomposed into a linear combination of cosine functions, and all the summation operations needed are performed by the SI method. The complexity of the proposed method is only O(N) independently of the kernel size. Experimental results show that the proposed method runs more than 200 times faster than the standard NLM and still retains similar performance. The proposed method is also evaluated with synthetic and real synthetic aperture radar data. The filtered images and quantitative measures show that the speckle is well removed while edges and shapes are preserved.