Image Smoothing Algorithm Based On Local Structure Information Research

Natural images contain rich and varied contents,and people often obtain effective information by analyzing images.However,the image will inevitably be interfered by various noise information in the process of imaging and transmission,resulting in the subsequent image processing tasks（such as image edge extraction,image recognition,etc.）will increase a certain degree of difficulty in implementation.The main task of image smoothing is to retain the main structural information while filtering the unimportant details in the image.Therefore,as a preprocessing process in the field of image processing,image edge smoothing technology also occupies a key position.The data items in the existing global optimization algorithm will carry out fidelity for pixels with gradient（including structural pixels and non-structural pixels）,so the texture information filtering effect is not obvious.In addition,it is found through the research that the penalty items of the current algorithms all use different patterns to constrain the gradient of the output image,and control the smoothing effect of the image by adjusting the smoothing parameters,which will also lead to the algorithm seeking a compromise between the smoothing ability and the retention ability of the weak gradient edge information.By analyzing the above problems presented by the global optimization algorithm,the main research content of this thesis is as follows:In order to avoid the problem that the weak gradient edge information is smoothed and the noise texture information is retained due to the strong gradient characteristics similar to the structure pixels,a global optimization algorithm framework is proposed.By calculating the similarity between the current pixel in the image and the pixel in the local neighborhood,the method initially determines whether the pixel to be processed is located in the structure area.Based on this,the two weights after the sub-item are designed to constrain the pixel to be processed to be similar to the original image or similar to the filtering result,so as to realize the retention of structural information and the smoothing of unstructured information.Due to its strong flexibility,the framework is suitable for most global optimization algorithms.Finally,the framework is applied to the smoothing algorithm example based on₂,₀,₁ and to conduct comprehensive comparison experiments with other algorithms,and PSNR and SSIM evaluation indexes are used to evaluate each algorithm on the BSD300 data set,which verifies the practicality of the proposed framework,and the improved algorithm can not only keep the weak gradient structure information,but also maintain the structure information.The image smoothing effect is better than the original algorithm.In view of the penalty terms in the global optimization function,image smoothing is realized by using different norms to constrain the output image gradient.Sometimes,in order to obtain better smoothing effect,some weak structure information is sacrificed,resulting in color distortion and detail blurring in the output image.In this thesis,a structure preserving image smoothing algorithm based on Locally Linear Embedding（LLE）is proposed.Combining with the idea of LLE,this method uses the relationship between pixels in the local area of the image,and uses the₂ norm to constrain local similarity to achieve image smoothing.Combined with the method of reweighting data items proposed above,the validity of reweighting data items is further verified from both theoretical and experimental aspects.Finally,the effectiveness of the proposed algorithm in preserving edge smoothing is verified by a series of comparative experiments,and it can effectively avoid artifacts,color distortion and block phenomenon.