Study On Several Problems In Deformable Shape Analysis And Recognition

The related research of computer vision, image understanding aims to make themachine “see” the world. The research is the interdisciplinary of neurobiology, signalprocessing, image processing, machine learning, and is an important area of artificialintelligence. Theassociatedtechnologicalachievementscangreatlychangetheproduc-tion of human life, which can be widely used in aerospace, industrial and agriculturalproduction, militarydefense, household, healthcare, transportation, navigationandoth-er aspects.Shape is an advanced visual feature, which is one of the most important research-es in the field of computer vision, image understanding. The most important featureof the shape is the deformation characteristics. In order to solve the difficulty defor-mation matching problem, in this paper, two levels start from the shape topologicalcharacteristics and the local sub-shape. The corresponding relationship, classificationand similarity of the set of points about the deformed shape are carried out in-depthresearch. We present the global topology information measure algorithm, the algorithmof calculating the contour length of shape in the gray-scale image, the shape classifica-tion method based on the sub-shape and the similarity measure algorithm based on thesub-shape.The main works of this paper are as follows:1. We propose a novel corresponding relationship matching method for the set ofpoints, which is priority considering global topological property of the set of points.The traditional point set matching methods are difficult to deal with consistent andlarge deformation shape. To solve this problem, we first transform the shape into agraph, and then propose the concept,“weighted pseudo common path”, which can ef-fectively measure the global topology similarity of the two shapes; On this basis, we analysis the traditional point set matching algorithm based on the optimization of theenergy function, and summarize the conclusion that the point set matching explicitlyor implicitly containing a set of points of space transformation and local adjacency in-formation; And thus design an energy function balanced global topological similarityand local shape similarity, which can be used the available dynamic programming tosolve. The energy function can not only find the corresponding relationship of pointsset of the deformed shape, but also can be able to estimate the similarity except theglobal deformation influence. Comparing to the classic point set matching algorithm,the experimental results show that the proposed model in this paper can consider theglobal topological property of point set, which can maintain the global topology of thedeformable shape invariance in the matching process, thus it will not failure in the largedeformation case.2. A contour length estimation algorithm for shape is proposed based on the cubicspline interpolation and the pixel gray information.First using cubic spline fits the local contour gray silhouette image edge, and thearc length of the spline function is approximated by the Taylor series. The whole cal-culation process avoids the specific parameters of the calculated spline fitting shapecontour, and only use four columns around the estimated position and the pixel valueof certain row. The method has linear computational complexity. The theoretical er-ror analysis shows that, in different resolutions, the relative error ceiling exponentiallyreduced with the reduction of the pixel convergence quantization step. The experimen-tal result shows that, comparing with the classical algorithm, the relative error on theclassic artificial data sets is lower, and has the same order of magnitude relative erroron the simulation of the actual data set and on the manual data. And with respect tothe classic B-spline estimation algorithm, the algorithm with the improvement of imageresolution, the relative error stable reduced.3. Proposed a shape Classification algorithm based on sub-shape.For deformation shape classification, application sub-shape method describing thedeformable shape or class can get rid of containing the effects of shielding shape. Wepropose sub-shape description based on optimal public distance. This algorithm canbe used not only shape matching, but also can be combined with other image point set feature recognition or matching images; Propose the definition of “request shapeclasses”; Proposethenotionof“separationabilityof2classes”withMutualInformationwhich can distinguish shapes to2classes; And a shape classification algorithm basedon the sub-shape class is presented. The algorithm improves the retrieval rate of theclassic data set on the deformed shape class.4. A Similarity algorithm based on sub-shape shape is proposed.For the sub-shapes can fullyreflect the invariance characteristics of the local shapeofthedeformedshape, thesimilaritymeasurealgorithmbasedonthecontoursub-shapeis given in this paper. For the local shape of the joint deformation, the shape of the sub-region-basedsimilaritymeasurementalgorithmispresented. Andweproposealgorithmcombinedwiththeshapeofthetwosub-metrics: toreducetheimpactofthedeformationof shape similarity, in the contour and regional, measure the shapes of the part of thealgorithmcorrespondingtotheapplicationofmeasurementalgorithm,respectively. Thealgorithm can significantly improve the retrieval rate on the classic data set of deformedshape.