Contour Information Based Image Pattern Matching

As an important research topic in computer vision,contour information based image pattern matching is widely used in industry,agriculture,commerce and daily life.The existing methods can be divided into two types: bottom-up approach and top-down approach.Among them,the representative method of the bottom-up approach is Generalized Hough Transform(GHT),which transforms each contour point from plane space to parameter space,and completes the matching between the template image and the object image by the way of voting and counting.The representative method of the top-down approach is Active Contour Model(ACM),which completes image pattern matching by evolving the initial contour toward the object edge.The evolving process is accomplished by minimizing the energy functional containing a priori knowledge of the shape.There are some problems to be solved in these two methods.Among them,GHT lacks the ability of matching the object with nonlinear deformation;while ACM needs an additional shape term containing a priori knowledge of the shape in its energy functional,which increases the complexity and abstractness of numerical solution,and lacks the controllability and flexibility.In this paper,we focus on solving these two problems.In order to improve GHT's ability of matching the object with nonlinear deformation,we extract the local topological structure features in the neighborhood of contour points when building reference table,and use feature matching method in voting process.In order to avoid the complex and abstract numerical solution brought by the shape term in energy functional,we propose a novel engineered matching framework,which substitutes the shape term with the mechanism of shape-space projection.Furthermore,we improve the controllability and flexibility of contour evolution by driving the sampling points of contour with optimal steps.The main research results and innovation of this paper are summarized as follows:(1)In GHT based image pattern matching,we propose the local topological structure features of the contour point neighborhood,which can describe the spatial distribution characteristics of object contour.These features consist of Generalized Crossing Number(GCN),Neighborhood Edge Direction Angle(NEDA),Angle from the Center of Gravity(ACG)and Distance from the Center of Gravity(DCG).Among them,GCN can describe the regional complexity of contours.NEDA can describe the regional directional feature of contours.ACG and DCG can record the feature of relative position between contour point and reference point.The comparative experiments on Chinese handwriting identification show that our method improves GHT's ability of matching the object with nonlinear deformation,and achieves better identification performance than current similar methods when keywords extraction is employed.(2)In ACM based image pattern matching method,we incorporate Shape-space Projection with B-spline representation of contour,Active Shape Model,CV model based driving force and a variety of contour evolving strategies,forming a matching framework with controllability and intuitiveness.The two shape-spaces proposed in this paper,Extended Affine Transformation Shape-space and Point Distribution Model Shape-space,can confine the arbitrary deformation produced by contour evolving process into the range of linear deformation and nonlinear deformation of same-class objects,respectively.The projections on the two shape-space can also determine the optimal value of parameters in shape model,and avoid the complex and abstract numerical solution with the same performance as the shape term in energy functional.Furthermore,our method has the characteristics of controllability and flexibility since it drives each sampling point of contour with an optimal step size,and can conveniently achieve the specific purpose of contour evolution.(3)The three contour evolving strategies proposed in this paper,Fast Contour Evolution(FCE),Contour Selection Evolution(CSE)and Sub-shape-space Selection(SSSS),can improve the speed and matching accuracy of curve evolution.Among them,FCE calculates the optimal step size of each sampling point driven by the force of CV model,improves the speed of evolution and makes the contour converge to the object edge more accurately.CSE classifies the sampling points according to the distance between these sampling points and object contour,and drives only the sampling point satisfying the distance condition.It solves the inaccurate object contour matching problem under the occlusion condition,and improves the robustness of the matching algorithm.SSSS provides some candidate means to constrain linear deformation more finely,and can effectively prevent mismatch and overmatching in evolving process.In addition,the preprocessing method proposed in this paper,Intensity Combination Method,transforms the test image in grayscale form into a sequence of binary images,decreasing the complexity of image pattern matching.The comparative experiments on synthetic images,road traffic sign dataset,MNIST dataset and ETHZ dataset show our method outperforms similar current methods in term of matching accuracy.