Abstract
Abstract. In this paper we propose automatic image denoising method based on Hermite functions (HeNLM). It is an extension of non-local means (NLM) algorithm. Differences between small image blocks (patches) are replaced by differences between feature vectors thus reducing computational complexity. The features are calculated in coordinate system connected with image gradient and are invariant to patch rotation. HeNLM method depends on the parameter that controls filtering strength. To chose automatically this parameter we use a no-reference denoising quality assessment method. It is based on Hessian matrix analysis. We compare the proposed method with full-reference methods using PSNR metrics, SSIM metrics, and its modifications MSSIM and CMSC. Image databases TID, DRIVE, BSD, and a set of dermatological immunofluorescence microscopy images were used for the tests. It was found that more perceptual CMSC and MSSIM metrics give worse correspondence than SSIM and PSNR to the results of information preservation by the non-reference image denoising.
Highlights
Use of the self-similarity is one of the main classical ideas in image denoising methods
This can lead to poor noise reduction along edges where the gradient has a different direction in each pixel of the edge This shortcoming is overcomed by LJNLM-LR by rotation of components of the feature vector to the coordinate system aligned by the image gradient
We can see that the parameter selected using SSIM is the closest to the parameter found by the non-reference method
Summary
Use of the self-similarity is one of the main classical ideas in image denoising methods. To overcome this shortcoming of NLM several methods has been proposed including LJNLM-LR (Manzanera, 2010) and GFNLM (Wang et al, 2012) In these methods weights depend on the Euclidean distance between the feature vectors which characterize the patches. In GFNLM features are based on Gabor functions Another shortcoming of NLM is that the method does not consider rotation of blocks i.e. pixels lying on one edge, but with different gradient directions, will be considered different and have small weights. A method for automatic parameter selection for image denoising algorithms has been proposed in (Zhu , Milanfar, 2010) It uses structure tensor analysis with fixed scale derivatives estimation.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
