Abstract: Aiming at the problems that using regular square patches for denoising might cause artifacts, or when the image suffered from af
fine deformation, the number of similar patches found by square patches would be reduced, we proposed a BM3D (block matching and 3D filtering) image denoising algorithm based on an affine invariant patch similarity measure. Firstly, in the process of block matching, elliptical patches were used to replace regular square patches for grouping by block matching to form three-dimensional elliptical patch groups. Secondly, an affine invariant patch similarity measure was used to calculate the distance between two elliptical patches and judge their similarity. This measure performed a patch comparison by automatically adjusting the size and shape of the patches, and more similar patches could be found and used for image denoising. Experimental results show that the algorithm can effectively improve the denoising effect and obtain good visual effects. Compared with the original BM3D algorithm, 
this algorithm has better peak signal-to-noise ratio.

Key words:  , image denoising, structure tensor, affine invariant patch similarity measure