Multiscale Image Processing Technology Based On The Evolution Approach

Image evolution model is one of relatively active fields of image processing,widely used in image smoothing, denoising, segmentation, feature extraction andmatching, etc. Researching on this field pays an important theoretical anddemonstration significance in the field of image processing for the establishmentof a strict mathematical system and its application.Based on large amount of research literature, we analyze the theoreticalrelationship between different images' evolution models, and establish a unifiedtheoretical framework, and then study the mechanism of the evolution model andassociated numerical algorithms. After that, we start to study Gauss model whichis relatively simple; by using singularity theory, the evolution characristic ofmodel feature points is researched, and the deep structure of Gauss model isrevealed theoretically. Based on the research of general image feature pointstheroy, this paper focuses on two types of feature points: the bifurcation point andthe curvature zero-crossing point, and applies the model and algorithms on imagematch and edge detection. Simultaneously, in the field of another important imagefeature—skeleton research, in this paper we propose a new method for imageskeleton extraction based on multi-scale learning theory from the idea of level setbased on energy function. In addition, since the Gauss model is a class of isotropyoperator, we will construct a new anisotropic diffusion equation model used inimage denoising, so that it has better adaptability. The main research and theinnovative work are as follow:(1) By improving the analysis method of the image structural similarity, thepaper constructs a new nonlinear diffusion adaptive equation to be used in imagedenoising. The corresponding algorithm can adaptively determine the number ofiteration steps and preserve image edges at the same time. Compared with thetraditional anisotropic model, the algorithm is simple. It can effectively removeimage noise without the manual intervention while preserves image edgeinformation better. So it has better adaptability.(2) Demonstrating the evolution of two-dimensional scale image featurepoint using singularity theory, especially the internal mechanism of the creationand the annihilation of feature points, this paper gives the strict description andthe demonstration of the deep structure of image. These studies can provide a theoretical for current research using the scale-space, and the nature evidence andanalysis of the feature points in the paper provide some algorithm basis forapplying the scale space to image match,motion tracking,segmentation etc.(3) The paper analyzes two kinds of important image feature points: thefundamental nature of the bifurcation point and the curvature zero-crossing point,and on this basis, we give some new image matching algorithms and edgedetection algorithms. The research about the two kinds of feature points providesa good attempt for studying image form the multi-scale perspective, and offers apractical idea for broadening further the multi-scale applications.(4) Through improved fast marching method, presents an algorithm ofextracting the narrow band image skeleton. The algorithm can get a continuoussingle-pixel skeleton, and effectively overcome the intermittent and burr problemsappearing in the current skeleton research. Because the algorithm has good noiseimmunity, at the fields of medical image processing and medical diagnosis, it laysa good foundation.The studies above systematically depict typical geometric evolution nature ofthe multi-scale image, and use them in the field of image matching,detection andso on, which provides new ideas and theoretical basis for further integratingin-depth study of multi-scale image and expanding the application field. Inaddition, this study builds on the unified framework of the evolution idea, whichis conducive to grasp the common of the different research models currently formoverall; and then lay the foundation for constructing a suitable optimizationalgorithm by analyzing the intrinsic link between the model and mechanism andlearning form each other.