Shape Recognition And Image Segmentation Method Study

Shape is recognized as one of the most fundamental feature used to describe image content.Shape classification and retrieval plays a key role in computer vision and image analysis.Related methods have been broadly applied to target detection,image retrieval, and object recognition.However,to extract and describe shape from an image is still a difficult task.This is because when a 3D real-world object is projected onto a 2D image plane,one dimension of object information is unavoidably lost.To make the problem even more complex,shape is often corrupted with noise,defects,occlusion and arbitrary distortion.Therefore,developing more effective technologies to deal with this problem remains in high demand.This paper first proposes a novel feature-based invariant descriptor called Radon Composite Features(RCF) to represent and identify shapes.Instead of normalization and direct analysis in the spatial domain,this algorithm uses Radon transform to parameterize the generalized morphological properties of groups of shapes.A modified Radon transform is proposed in order to make the transform matrix invariant to the scaling of the shape.To extract the shape information encoded in the Radon transformation plane, both spectral and structural means are proposed.Fourier coefficients and three structural features which have strong descriptive abilities are extracted.The proposed method overcomes the drawbacks of existing shape representation techniques since it accomplishes the invariants to common geometrical transformations without any normalization process, which usually causes inaccuracies.A novel hierarchical strategy with RCF can achieve low complexity and coarse-to-fine retrieval,and perform accurately when retrieving shapes, while remaining robustness under variations.Experiments demonstrate that compared with some state-of-the-art approaches,RCF provides a higher degree of discrimination. The proposed method has also been successfully applied to SAR image classification.To extract an object and acquire its shape from a give image,an image segmentation stage is necessary beforehand.Deriving from the Artificial Life(Alife) theory,this paper then proposes an Artificial Co-evolving Tribes(ACT) model and applied it to solve the image segmentation problem.In this model a given image can be seen as a living environment. Unit that initially resides on each pixel of the image is considered as a living individual.There is interaction among them during the co-evolving process.Each agent intends to find its congeners according to both spatial and feature distances,immigrate to the area with suit best for it,and alter itself according to the local environment.The individuals in this model making up the tribes effect communication cooperatively from one agent to the other in order to increase the homogeneity of the ensemble of the im age regions they represent.Two remarkable properties,that is,the monotone contraction and the conservation of the system are proved.Stability and scale control of the proposed method are carefully analyzed.Experimental results are presented and compared with two related segmentation methods,both quantitatively and visually.This paper also includes the discussions of the results matching with human visual perception.Being viewed as a novel application for self-organized complex systems,the proposed approach breaks a new path to the treasure trove for image segmentation.Pattern classification is the stage after the features are extracted from shapes.Chosen the proper classification method usually can obtain better recognition results.In this paper,a novel classification algorithm using Local Probabilistic Centers(LPC) is proposed.This method works through reducing the error-prone samples and restricting their influence regions.Especially when the data distributions are overlapping,or when the samples are deeply polluted by noise,traditional k-th Nearest Neighbor(k-NN) algorithm will be severely influenced and generate poor results.Local Probabilistic Centers(LPC) approach employs some prototype-based technology to reduce the outliers and shrink their influence regions.Two different measuring methods are also proposed, one is the distance between query and local probabilistic centers,and the other is the computed posterior probability difference of query and the nearest categorical center.Although both measures are effective,related experiments show the second one achieves the smaller classification error.This paper also investigates the expectation risk and stability of the proposed method.A set of experiments are conducted on both constructed datasets and real world datasets.The classification results are carefully evaluated and compared with some related methods,which demonstrate that LPC successfully avoids the drawbacks of the traditional k-th Nearest Neighbor algorithm and substantially improves the classification performance.