Font Size: a A A

Research On Extraction And Application Of Pattern Features

Posted on:2008-01-26Degree:DoctorType:Dissertation
Country:ChinaCandidate:S F XieFull Text:PDF
GTID:1118360272466898Subject:Control Science and Engineering
Abstract/Summary:PDF Full Text Request
Features Extraction plays an important role in image processing and pattern recognition that are widely applied in many fields. The main content of this thesis includes four aspects: Extraction of image invariant feature; Detection of image edge feature; Identifying and cleaning of the terrain elements based on feature in scanned topographic maps; Extraction of the statistical characteristics of bubble-phases in a fluidized bed boiler. The main tools and the methods that this thesis applies include Polar coordinate transformation, Radon transform, Walsh transform, Fourier transform and Wavelet analysis.The main contributions of this thesis are given below:Two methods for extracting image invariant features are presented.Firstly, by use of Polar coordinate transformation, Walsh transform and Fourier transform, a Walsh-Fourier moments that are rotational invariant are presented, and the methods for solving the scaling and translation invariance problems are also given; Secondly, by use of the linear Radon transform and Fourier transform, a Radon-Fourier invariants that are rotational and translation invariant are presented, and the method for solving the scaling invariance problems is also given. Experimental results show that these two types of invariant features are very strong to describe image characters, and a higher recognition rate can be obtained. Especially, in the anti-noise performance, they are obviously better than the classical Hu invariant moments and Zernike invariant moments. Two methods for image edge detection are presented.The first method is based on local radial projection curve. It includes two algorithms. Firstly, a class of ring operator corresponding to local radial projection curve is given, and then the step-edges are detected by cyclic convolution between the local radial projection curve and the ring operator; Secondly, the Fourier transform coefficientes of the local radial projection curve are used to detecte edges, that is, first-order and second-order harmonic coefficients are used to detecte step-edges and roof-edges respectively, and the DC component coefficient is used to determine the types of roof-edges. Experimental results show that the step-edge detection algorithms presented are better than the classical Soble operator and Laplacian of Gaussian operator, and is a match for the optimal Canny operator; The roof-edge detection algorithm presented is obviously better than the algorithms based on zero-crossing of the first order differential and main curvature.The second method is based on the local pseudo-linear Radon transform. Firstly, for several columns adjacent to the current column of image, the pseudo-linear Radon transform matrix is obtained, and then the differences along each column of the transform matrix are calculates; Secondly, all the step-edge points on the current column of image are detected by the absolute values of the differences those are greater than a given threshold. Experimental results show that the method presented is better than the classical Soble operator and Laplacian of Gaussian operator in the anti-noise performance, and is better than Canny operator in the edge details and precise.Two pre-processing technologies for vectorizing scanned topographic maps are presented.The first technology is to segmente and recognizes the elevation value in the scanned topographic maps. Firstly, an algorithm for extraction of connected components based on mathematical morphology is employed to find out elevation value notes; Secondly, two methods are used to recognize the elevation values. The one method is based on the multi-scale fractal dimension features of the ring projection curve of the digital character; The other method is that the rotation angle of digital characters in an elevation value are calculated by the smallest covage circle and linear fitting, and then the elevation value was recognized by use of template matching method. Experimental results show that the approach can accurately segmente and recognize the elevation values.The second technology is to identify and clean the non-contour elements in the scanned topographic maps. All the non-contour elements are divided into four types according to various characteristics, and then four methods for identifying and cleaning different types are presented respectively. These methods include: Extracting connected components for identifying the independent elements; Local horizontal and vertical projection for identifying massive elements; Local ring projection for identifying circle elements; Local radial projection for identifying line elements those are non-contour. Experimental results show that these methods can effectively identify and remove most non-contour elements. It is very helpful for contour vectorizing. The method for extracting the statistical characteristics of the bubble-phases in a fluidized bed boiler is presented.Firstly, according to the differences of amplitude and spectrum characteristics between the bubble-phase signal and dense-phase signal in a fluidized bed boiler, the points alternated from one phase to another phase can be determined exactly by using wavelet method, and the bubble-phase signal can be picked up; Secondly, a loca1 cross-correlation method is used to calculate bobble movement parameters, and the distribution law of the bubble-phases can be obtained. Experimental results show that the bubble sizes and velocities are both obeyed log-normal distribution. It is very helpful for designing and automatic control of fluidized bed boiler.
Keywords/Search Tags:Feature extraction, Edge detection, Polar coordinate transformation, Radon transform, Scanned topographic maps, Elevation values recognizing, Fluidized bed boiler, Bubble-phase
PDF Full Text Request
Related items