Squama Cancer Cell Recognition Based On Moment Features

With the development of computer technology, the use of image recognition become more and more widely in iatrology. In the field of automatic recognition of cell, although there exit plenty of recognition algorithm, no algorithm is perfect now. It is because of the varieties and complexity of cells.In this thesis, the image segmentation is performed by Support Vector Machine, the image moments are used as cell image features and the Support Vector Machine is used as classified tool and squama cancerous cells are used as recognition objects for establishment of a general algorithm for various kinds of cells.Moments are used as the features of image for its multi-distorted invariant. The moments that are taken as features can keep invariable when the images are rotated, scaled and shifted. Hu first utilized moments as image features in 1961, and he proposed a set of regular geometric moments with a multi-distorted invariant as image features and it is very difficult to be used to recovery an image with geometric moments. Then the rotational moments, complex moments, Legendre moments, Zernike moments are proposed in the field of image moments, the Zernike moments have the best description for image and the antinoise is strongest among all kinds of image moments,. Because the Zernike moments perform bad for description of small image, Sheng yun-long et al introduced the orthogonal Fourier-Mellin moments in 1994, they have the similar properties as Zernike moments when they are used to describe a big image, but they have better properties when they are use to describe a small image. Later Ping zi-liang et al proposed Chebyshev-Fourier moments with the similar properties as that of the orthogonal Fourier-Mellin moments. Then Ping zi-liang et al proposed Jacobi-Fourier moments and point out that all the moments constructed by polynomials such as Legend moments, Zernike moments, Chebyshev-fourier moments and the orthogonal Fourier-Mellin moments is just a special case of Jacobi-Fourier moments, they can be derived from the Jacobi-Fourier moment with different parameters. In 2003, Ren hai-ping proposed the image moments with the best properties:radial harmonic Fourier moments. In this paper, the radial harmonic Fourier moments will be researched deeper, a generic form of radial harmonic Fourier moments will be proposed:Exponential-Fourier moments, Exponential-Fourier moments have the better properties than that of the radial harmonic moments constructed by the triangular function.There exists a problem in the computation of moments with multi-distorted invariants:all the image moments are defined and calculated in polar coordinate system in order to keep the multi-distorted invariants of images, but all the image are saved in the Cartesian coordinate system in computers. For the computation of image moments, the images have to be converted into discrete polar coordinate system. And the image reconstruction is performed in polar coordinate system, it must be converted into the Cartesian coordinate system. The conversion not only increases the computational load and also results in interpolation error. In the thesis, the calculation methods of moments with multi-distorted invariants and images reconstruction in Cartesian coordinate system are discussed.After extraction of image features, the optimal classification surface can be found in the feature space by the SVM, and then, a general algorithm about cell recognition is obtained by using the optimal surface to classify the images. The every part of the algorithm is general, such as cell segmentation, feature extraction and cell recognition.During the process of research about the cell image recognition, some developments are made:1. The method of image reconstruction in Cartesian coordinate system is proposed. All the known methods of image reconstruction with moments is in the polar coordinate system, the reconstruction image must be converted into the Cartesian coordinate system. In this paper, the method of image reconstruction in Cartesian coordinate system is proposed in the third chapter. The method avoids the coordinate conversion so that it eliminates the rounding error and shorter the calculation time.2. Among the various kinds of image moments, the Radial Harmonic Fourier moments that are proposed by Hai-Ping Ren have the best performance. By research of the RHFMs, we found that the RHFMs proposed by Ren Hai-Ping is just a special case, there exits many RHFMs with different forms and same performance.3. A new image moment with best performances is proposed:Exponential Fourier moments, it takes shorter time for its calculation and has same performance as that of RHFMs, but it has simpler form. It is the best image moments so far.4. The algorithm of calculation about moments with rotation invariant property in the Cartesian coordinate system. All the known methods for image moments calculation are in the polar coordinate system in order to keep their rotation properties. But imges are shown in Cartesian coordinate system, the image coordinates are need to be converted. It must be lead to rounding error, and conversion of pixel coordinates will increase computational load. In this thesis, a algorithm of calculating the moments in Cartesian coordinate system are proposed in the fifth chapter, it is unnecessary to convert the pixel coordinate during the moments calculation, it not only speeds up the image moments calculation but also eliminates the conversion error about coordinates.5. An automatic recognition system is proposed by means of Exponential-Fourier moments and SVM. Various kinds of extraction algorithm and classified method are necessary for different cell due to the cell diversity. But the algorithm for the calculation of Exponential moments are same, so that we can design a general algorithm which take the image moments as image features for different kinds of cell.