Multi-scale Inhernet Variation Based Structure Detection And Texture Filtering Algorithm Research

With the development of machine learning and artificial intelligence,image as its data source,people's image processing requirements are also getting higher and higher.Image filtering in the past is mainly to remove noise,but in recent years a new view was put forward that image filtering should not only removes noise,but also filters out unnecessary texture details,which is called texture filtering.This processing has powerful applications in detail enhancement,image abstraction,region segmentation,object extraction,inverse half-tone and so on.In texture filtering,the core is the separation of structure and texture,but they are similar.How to effectively keep the structure while filtering out the texture is a difficult point in the process of texture filter.However,most of the existing filtering algorithms tend to deal with weak gradient texture images,if the texture gradient is large,the algorithm will fail.Some algorithms based on global optimization both keep structure and image smoothing in the energy function,but the parameters are difficult to adjust and can not guarantee the effect.In order to ensure that these two do not interfere with each other,we consider to deal with the structure and the texture separately and detect structure before texture filtering.However,the existing structure detection algorithms are not well suited for texture filtering algorithms,so we first propose a structure detection method based on multi-scale inner variation,and then propose a structure-guided texture filtering algorithm.This algorithm can solve the limitations of the existing algorithms with strong gradient texture.The main innovations of this paper include the following three parts:Firstly,we propose a multi-scale inherent variation to distinguish structures and textures and this method is more reliable than single scale ones,regardless of the gradient strength and scale size of the texture.And then we extract six discriminating features according to the multi-scale inherent variation as training characteristics.Secondly,in view of the shortcomings of the SVM classification results,we design three step post-processing work,including outlier rejection,multi-scale breakpoint connection,and curvature-based structure correction,and solve the burr,missed inspection,structure offset respectively and finally get a fine structure detection map.Thirdly,we propose an adaptive texture filtering algorithm under the guidance of structure,which solve the problem of halo and color cast of local filtering.The filtering result can keep the structure information and inhibit the texture detail efficiently,and the effect of the improvement is obvious.