Morphological Analysis Method And Its Applications In Image Feature Extraction

Mathematical morphology is an effective way of image processing and analysis in the spa-tial domain.It has lower computational complexity due to calculation by convolution operation.In addition,the morphological operator is convenient of implementation by using hardware,therefore it has the characteristics of fast and efficient in practice,especially for a large amount of image data processing and analysis.Image feature extraction is one of the basic problems in image analysis,and also a key step from image processing to image understanding.Thus it provides an important basis for image registration in computer vision and reconstruction,image classification and object recognition as well as image retrieval.Therefore,it has the important theoretical and practical significance to establish a image feature extraction method with high accuracy and robustness.The purpose of this paper is to establish advanced morphology operators to supplement the drawbacks of existing operators and enrich theoretical system of morphology.Then,according to the characteristics of images(gray image and color image)and the type of feature,proper morphological analysis methods are constructed for feature extraction in practical in order to test the effectiveness of the proposed morphological models.Based on the introduction of the basic theory of morphology operators and feature extrac-tion,this paper lucubrates on the various morphology operators and morphology-based feature extraction methods with the help of combination of fuzzy sets,tensor model,manifold learning and soft computing techniques according to the type of images and features.The mathematical morphology analysis methods and applications in image feature extraction are discussed from the following several aspects:(1)In order to solve robustness problem of the gray-level morphology operator,based on fuzzy hierarchical translation and fuzzy inclusion,the variable precision fuzzy hit-or-miss trans-form model has been constructed by combining with variable precision idea and order statistical operator.Relying on the introduction of the concept of fuzzy translation,the method is able to not only relax the restriction of structuring elements in the gray-level,but also increase the ro-bustness of definition of structuring elements.Moreover,the proposed method can well reflect the matching relationship between the image target and the given structuring elements.For the robustness of gray-level image feature extraction,the paper proposed a variable precision method to suppress the influence of gaussian noise and image fuzzy on feature extraction.Then,based on order statistics idea the rank and soft variable precision fuzzy hit-or-miss operator is introduced to eliminate impulse noise in images.The rank and soft variable precision fuzzy hit-or-miss transform are respectively used to detection edge and corner by analyzing the neighborhood model of interesting points.The experiments are conducted on several noisy gray-level images in different levels to verify the availability of the proposed operators in interesting point feature extraction in noisy environ-ment.In addition,it also gives a discussion on the determination of algorithm's parameters in-cluding precision coefficient,rank and soft order.Finally,the usage of variable precision fuzzy hit-or-miss transform in image object detection is studied.The experiments of object detection are respectively conducted on strong noisy images,radar images,and medical images.Quali-tative analysis in theory and practice is presented to show the advantage of the proposed model compared with existing robust gray-level hit-or-miss transforms.(2)Aim at morphology analysis of color images,three kinds of tensor models are pro-posed in the paper.First,the color tensor model is defined by incorporated physical meaning of the vector in HSI color space with the spectrum decomposition of the second order tensor.Second,the texture representation model based on generalized structure tensor is established by constructing texture mapping of the color image in the view of differential geometry theory.Finally,the paper puts forward to a second-order three-dimension tensor called the mix tensor model recurring to the corresponding relationship between tensor spectrum and color and tex-ture information.In order to establish the tensor morphological operator models,two tensor partial orderings are constructed based on the marginal ordering of tensor spectrum and tensor distance measure respectively.And on this basis the relevant tensor morphological operators are defined.In accordance with the need of application in feature extraction and the characteristic of different models,color tensor-based morphological gradient is considered as an edge extrac-tion operator and the qualitative analysis is carried on performance with respect to different tensor distance measure.For texture image edge extraction problem,both texture tensor-based morphological dilation and mix tensor-based morphological gradient are deliberated.The per-formance and applicable scope are given through the experimental analysis.(3)Dealing with the problem of color image segmentation,the paper aims to study the irregular regional feature method for color images.In the framework of granular computing,the modified color morphological jump connected operator is adopted to divide an image into several homogeneous irregular connected regions taken as basic granules.Based on the regular-ization of basic granules morphological reconstruction is used to map all objects into the same Euclidean space.Then,the revised Laplacian eigenmaps manifold learning method is applied to extract low-dimension manifold embedded into high-dimension Euclidean space as feature representation of irregular regions.Finally,Markov Chain Monte Carlo(MCMC)method is explored for fusion of basic granules to solve the problem of color image segmentation.The segmentation experiments by proposed approach are carried on several natural images from Berkeley standard image database.Then,the effects of algorithm's parameters on segmentation performance are researched.Quantitative analysis by three indexes shows the effectiveness of the algorithm compared with classical color image segmentation.