Shape Classification Based On Vector Field Consensus

Shape classification is very important in the field of computer vision. It is the basis ofdirections like object detection, text detection and shape evolution, and It has a brilliantfuture in the field of disease diagnosis, business retrieval, handwriting input, identifierrecognition in daily life. To the computer, shape is a kind of direct description way ofobject. We can analyze and extract information from shapes such as object class. Shapeclassification is to match and classify the shape in an image, and find out the class itbelongs. It is hard to deal with problems about the using of shape information, noise andlocal information, scale-invariant and rotation-invariant, deformation of object when wewant to do shape classification.This thesis summarizes current shape matching and classification algorithms. Theycould be divided into three kinds: shape matching and classification algorithms based oncontour, region and skeleton.Deep analysis on the advantages and disadvantages of matching algorithms andshape classification algorithms is given in this thesis. A novel method named “shapeclassification based on vector field consensus” is proposed, and it is based on the vectorfield consensus, matching algorithms and current shape matching and classificationalgorithms. First, points are sampled by the shape matching algorithm based on contour,and then the histograms of the points are created. Second, the similarity matrix iscomputed and the matching solution is computed by the Hungarian algorithm. Third, thewrong matchings are divided from the matchings by the vector field consensus. Last, thesimilarities between shapes are computed based on the matching solution, and shapematching and classification algorithm is done. It could distinguish wrong matchings fromright ones, and dig out the latent relations of matching points, thus it could greatly approvethe precision of matching. The experiment results proved the shape classification based on vector field consensus algorithm have a large progress compared with old classificationalgorithms.