Similarity Measure And Its Application In Image Non-local Filtering

Image denoising is the core and key work in the field of digital image processing, its fundamental and practical occupy the important position in the field of digital image and cannot be ignored. Noise not only affects the quality of the image itself, but also can bring the wrong information to disturb our judgment, and thus for noise removal is necessary. Non-local means filter opens up a new starting point for denoising, across the bondage of local, opened another door in the field of denoising. The previous denoising methods based on neighborhood pixels and the Non-local means filter breakthrough base on the way of blocks and blocks, by measuring the similarity of neighborhood block to estimate the degree of similarity between pixels and the pixels, through the weighted average of these points of finally filtering results, and obtain good denoising effect. However, the weighted Euclidean distance is susceptible to noise on the image, coupled with the Euclidean distance does not characterize the overall characteristics of the image block, similarity in measuring block would produce the phenomenon of mismatch,and it has seriously affected the denoised image visual effect.We designed an improved similarity measure nonlocal mean denoising method aim to balance neighborhood patches'similarity in the paper. Feature with an image block from the perspective of physics and applied mathematics evolved descriptors-normalized moment of inertia, it as part of the weight adjustment pixel similarity measure. The normalized moment of inertia is not only an image invariant feature descriptor of the translation and rotation, but also robustness to noise. By the different methods of similar block search experiment comparison analysis, NMI is more matching the image's local similarly. The proposed method has been evaluated on testing images with various levels white Gaussian noise, extensive experiments demonstrate that based on normalized moment'of inertia of nonlocal means denoising method achieves much better results than the classical NLM method not only the detail structure information, but also the peak sign-to-noise ratio and structure similarity have significantly improved. Especially, when the noise intensity increases, the algorithm can still keep good denoising effect. In addition, not just confined to the traditional experiment under white gaussian noise, also in different intensity of salt and speckle noise test, through experimental demonstrate the new algorithm of peak signal to noise ratio and structural similarity is improved, denoising effect is obvious.