Research On Novel Shape Analysis Techniques And Their Applications In Image Retrieval

Shape description and analysis techniques are the hot research topics in patternrecognition, computer vision and image understanding which have been widely applied inmany fields of scientific research and engineering technology, such as object recognition,image retrieval, image registration, product detection, biomedical engineering and etc. Themain work in this thesis is closely around the core problem of shape description and analysis.Based on the further research of the existing shape analysis techniques, several novel shapedescription and analysis approaches are proposed, and which are used for shape imageretrieval applications. The main work and contributions in this thesis include the followingaspects:(1) First, several shape compactness description methods are briefly introduced and thensome typical shape description and analysis methods proposed in recent years are presented.The construction processes of these methods are described in detail, their advantages anddisadvantages are discussed and also the possible improvement directions are looked forward.(2) Second, using the relative distribution of points’ position of the contour to describe ashape, a novel shape contour descriptor for shape description is proposed, named contourpoints distribution histogram (CPDH), under polar coordinate according to the thinking ofstatistics. This descriptor is not only satisfactory to the human’s visual feelings but also isvery simple to be calculated. And it essentially has the properties of invariant to scaling andtranslation. The dynamic programming algorithm is suggested to measure the distancebetween CPDHs, and that the dynamic programming algorithm can partly fulfill the needs ofthe CPDH’s invariant to rotation. The problem of the distance measurement between CPDHsis treated as the transportation problem in the operational research. According to thecharacteristics of CPDH an effective ground distance calculation method is suggested and thefinal distance between two CPDHs is obtained through the combination of the shift matchingand mirror matching in the process of distance easement. By a great deal experiments inseveral common shape databases, it is shown that the proposed algorithms, used in imageretrieval of shape with a single closed contour, can get favorable results.(3) Third, the shape region descriptors based on the moment invariants play an importantrole in pattern recognition and computer vision. According to the visual characteristic thathuman tend to discriminate different shapes more dependent on the contour parts of the object(i.e. the contour parts play the major role when people identifying different shapes), a kind ofcentral moment weighted method is put forward. This weighting method has the followingcharacteristics. When computing the image’s moment invariants, according to the differentdistance between the points and contour different weights are assigned to the points indifferent position. If the point in the shape is close to the boundary then it will has the biggerweight. And on the other hand, if the point in the shape is far away from the boundary then itwill has the smaller weight. The seven Hu moment invariants reconstructed based on theweighted central moment are selected as the image features. The reconstructed seven Hu moment invariants are still have properties of invariance to translation, scaling and rotation.By a great deal experiments in several common shape databases, it is shown that retrievalperformance based on the improved Hu moment invariants is much better than that based onthe traditional Hu moment invariants.(4) Fourth, A novel shape image retrieval algorithms based on compensation mechanismunder minimum circumscribed circle is proposed, which not only extracts the features of theobject area, and extracts the features of the background within the minimum circumscribedcircle area. Extracting the Hu moment invariants and Zernike moment invariants, commonlyused as features for describing the region, as image features. Similarity between two images isgiven by the Euclidean distance of the normalized moment invariants vectors. This method isnot only implemented simply, but compensates the human eye’s visual perception effectively.By a large number of experiments, it is shown that this approach can achieve better retrievalaccuracy and recall rates than that of an object area based only.(5) Fifth, Shape signature and Fourier descriptors are common techniques for shapedescription and they are widely used in pattern recognition and computer vision applications.A novel shape signature is proposed in this thesis, namely, multiscale contour flexibility shapesignature. After the discrete Fourier transform is performed on the multiscale contourflexibility shape signature, the Fourier descriptors are obtained. Contour flexibility basedFourier descriptor is a contour line function, which not only describes the whole deformationcharacteristics of the two dimensional shape profiles, but also reflects the local deformationcharacteristics of the contour sampling points. And it incorporates the global and localfeatures of the shape. Multiscale technique has solved the problem of elastic parameterselection and describes the shape features from coarse to fine. It is also easy to be calculated.By a great deal experiments in several common shape databases, it is shown that the bestretrieval results are achieved by the multiscale contour flexibility based Fourier descriptorcompared with other typical shape signatures based Fourier descriptors.