Research On Image Feature Extraction And Recognition About The Trace Space Projection

Image feature extraction is a key point in the study of the image content recognition,which involves research content of many disciplines such as image processing, patternrecognition, signal processing and machine learning. In study of image feature extraction,such features as geometrical invariability,the strong descriptive of image content and the lowinformation redundancy always draw people’s attention. In recent years, Radon domainanalysis and Trace transformation have become the latest extraction methods of imagefeatures. The former can achieve geometrical invariability of image features by Radontransformation and then by image features extraction in the transformation domain; the latter,which is a promotion of Radon transformation in two dimensional space, can transform thetranslation of image content, rotation and scale change into functional parameters change bydefining Invariant and Sensitive functionals. Then the image invariant features can beconstructed through the calculation of the parameters.The research mentioned above has obtained many achievements, but not yet has anydefinite model to analyze all these methods. Based on Radon domain analysis and Tracetransformation, a kind of “Trace space projection” image feature extraction method isproposed in this dissertation. The main research includes the following several aspects:1. Based on the further study of Radon domain analysis and Trace transformation theory,a kind of mathematical method called Trace space projection is proposed in this dissertation.By defining image trace line, a two dimensional image is projected into a three dimensionalspace. By intrgrating and transforming the three dimensional data space in different directions,this method can be used not only to detect image linear features, but also to extract thecharacteristics of image with geometrical invariability, which is of great application value.Theoretically,this method can unit Radon transform, Trace transform, Ridgelet transform andthe two dimensional Fourier transform into a same mathematical framework of Trace spaceprojection, and by which polar coordinate image moment method can be analysed. The Tracespace projection transformation formula, the extraction method of image features and theTrace space projection realization method is proposed in this dissertation, and general type ofTracespace projection is also discussed. 2. Concerning the fact that the face image with in-plane rotation by stochastic angle isdifficult to identify,a method by fusion of2DNPP and Trace space projection is proposed toextract and identify invariant features of face images rotation. First Trace space projection isdone to the image to extract a sinogram and double trace feature, then higher order spectralmethod is used to match and calculate the double curve, the characteristics obtained arediscovered of robustness for the in-plane rotation changes and the maintenance of the imageinformation by the bispectrum function. Finally2DNPP is used to reduce dimension andclassification to obtain the invariant feature. Compared with SIFT, Pseudo-Zernike momentmethods, the proposed method has better performance with higher recognition rate, and is ofstronger robustness to the white noise.3. The application of the Trace space projection in the image linear features extraction isfurther explored. A new T functional for SAR image Trace space projection data is proposedto extract the ship orbit characteristics and vehicle direction angle characteristics. Simulationresult shows that the method is effectively used in SAR image recognition. By doing twiceanalysis of Trace space projection of the checkerboard image, a kind of double trace spaceprojection method is proposed to obtain the checkerboard image parallel line positioninformation, and then through the geometrical relationship computation, the checkerboardimages Corner`s positions are got, This method can avoid noise interference outside thecheckerboard. In the image digital watermark embedding study, the linear feature position isobtained by Trace space projection for image analysing, and at the given place digitalwatermark embing, the watermark with transparency and robustness can be obtained.4. A novel image feature extraction method of fusing orthogonal polynomial moment andTrace space projection is proposed. In this method, the sinogram is obtained by calculatingTrace space projection. And the image sinogram is is decomposed in the orthogonal basisfunction space to obtain the low redundancy orthogonal characteristics of rotation and scalinginvariance. Compared with Zernike moment, Tchebichef moment and OrthogonalFourier-Mellin moment algorithm, this method is of better image shape and textureinformation description ability.At a given condition, the feature obtained by this method canalso be used to reconstruct the original image.The proposed Trace space projection image feature extraction algorithm can be used toextract the linear or the geometrical invariability characteristics of the image and unite Radondomain analysis, polar image moment analysis and many other methods into the sameframework for research, which is of great theoretical value. The Trace space projectionfeatures can also be applied in image feature extraction and pattern recognition, which is of certain practical value.