Non-local Mean Value Image De-noising Algorithm Based On Structure Information

With the fast development of society, all kinds of fields have growing needs of higher image qualities. However, digital images will be, during capture and transport, inevitably affected by various noises, which causes lower image qualities, and therefore the following processes, such as cut, recognition and feature extraction, could be seriously disturbed. Thus, it is vital to de-noise previously with the present capture and transport paths and the image qualities will be improved.Non-local mean value image de-noising algorithm is newly proposed. Massive redundant information belonging to the image itself is used to de-noise, which is the similarity between image blocks. When measuring the similarity between the blocks, non-local mean value image de-noising algorithm used Gaussian weighted Euclidean metric and only the geometrical structure information among image blocks is considered while the directional structure information of the block itself is not. Therefore, there are some errors when measuring the similar blocks, for example, finding wrong or missing similar blocks. Non-local mean value algorithm uses certain global filter parameters, which easily caused that details of images are too much smooth or smooth areas are not enough capable of de-noising. This paper has some improvements of non-local mean value algorithm. The improvement is as follows:(1)Local directional structure information similarity algorithm and the product Gaussian weighted Euclidean metrics are used as the measurement of similarity between image blocks, so that enhancement of neighbor directional structure information can improve the description of similarity between neighbors. Method of extraction of the information of image blocks is proposed, where similarity measurement is more precise after improved.(2) Canny factors are used to detect the margins of images, so that the areas of rich details and smooth areas can be distinguished. Therefore, the filter parameters would be adjusted according to structure complexity of the current searching neighbor. Different filter parameters are selected in different structural areas. In smooth areas, the filter parameters are large, while in areas of rich details small. The filter parameters can adjust themselves. The certain selection method of filter parameters comes up and gets validity by experiments.