Research Of Shape Features Statistics And Similarity In The Polar Coordinate System

With science and technology growing rapidly and internet technologyflourishing in each passing day, we come into a big data era with information’sexplosive growth. Especially visual information, as an important medium forinformation disseminating, How to make it fast and accurately becomes anincreasing problem. Nowadays, Traditional text-based image annotation has beenunable to cope with such a large number of image classification and retrieval,therefore, content-based image retrieval and semantic annotation becomes a newmeasure for current issues.Shape, as one of the image’s three base-level features (color, texture, andshape) and a high level visual feature, plays an important role in content-basedimage retrieval. Shape matching also becomes a hot frontier issue in the field ofimage analysis and computer vision. Although various complex shapes bring thestudy a lot of difficulties, many classical algorithms spring up with a large numberof scholars’ long-term continuous efforts.With the carefully review and analysis of previous work, combining chaincode’s ability for local contour describing and the hidden feature in centroidcontour distance, this thesis proposes a joint statistical descriptor of centroidcontour distance and chain code, a matrix feature, which is different fromtraditional feature vector.The main innovation of this thesis is as follows:Firstly, we present the concept of matching matrix. Matching matrix is notonly able to improve the statistical chain code’s alignment problems, but also torestore the shape from the flip transformation in the matching process.Experimental results based on MPEG-7database show that the retrieval rate is14%higher with matching matrix than minimum summation statistical directional code. Secondly, we present a novel shape representation and matching methodbased on joint statistics of centroid contour distance and chain code. According tothe joint analysis on centroid contour distance and chain code, the silhouette isdecomposed into several levels based on centroid contour distance. And then,analyze the frequency of each symbol in the chain code describing the silhouettelaying in each level to make the joint statistics of centroid contour distance andchain code, a matrix feature descriptor. Database experimental results show that thejoint statistics descriptor is not only better than a single feature descriptor, but alsobetter than combination algorithm with local and global by weighing method.This thesis applies statistical idea into shape feature analysis, and keeps theshape features to a great degree by using statistical idea instead of samplingmethod. And also, we propose a features joint analysis idea, which can enhance theshape representation ability of traditional shape descriptors. In the shape matchingstudy, we propose a matching matrix, a product of the concept that a literdimension leads to a directional alignment, which can restore the fliptransformation that other methods are not able to. We hope that people engaged inthis research area can get some inspiration from our ideas and algorithms in thisthesis, encouraging the development of this field together.