Research On Moment Method For Model Retrieval

As a feature extracting method, moments methods have important application value in image processing, computer vision, artificial intelligence, etc. Moment methods have many good characteristics such as compact, low computational cost, ease to be constructed geometric invariant, robustness to noise, ability to construct image, and so on, which makes moments methods occupy an important place in extracting shape features.Due to dimensionality problems and time consuming problem in extracting shape features, we compare the existing moment methods and analyse new moment methods defined in different space to extract shape features or construct geometric and blur invariant, for models retrieval.Based on the peers’ work, the contributions of this research work in this paper are summarized as follows:1In order to highlight the good property of Bessel-Fourier moment in describing shapes, a new moment named Bessel structure moment is defined and applied in image retrieval. The experiments show that the new moment inherits the characteristics of Bessel-Fourier moment and has better precision performance.2Due to moment geometric invariants, a new set of of Legendre moments in polar coordinate is introduced for image retrieval. The new approach is obtained by the two processes. Firstly algebra method is to find the rotated angle and scaled factor of original image. Secondly normal process is substituted to the polar equation to offset the rotated angle and scaled factor. Meanwhile, the features extracted by proposed method are invariant to translaion, rotation and scaling. Experiments verify the retrieval efficiency of the proposed method.3A new circularity measure method based on complex moment is proposed. And it is combined with gradient magnitude and Phase congruent by frequency structural similarity function for image retrieval. Experiments show that the proposed method has higher precision rate.4A novel approach called three shape feature functions is proposed based on Radon transform directly. The three corresponding shape features are defined as the symmetry of the shape, the width of contour and the fullness of the shape. They can reveal the information of visual and the inner structures of shape using characteristics of Radon transform. Theoretical and experimental results verify that the novel method invariant to rotation is more efficient in image retrieval and has better capability in handling a variety of shapes.5Based on the nonlinear Radon transform, nonlinear Radon moment is defined and conducted RST invariance and blur invariance. The retrieval efficiency and robustness of nonlinear Radon moment are different with the curves in Radon transform including ellipse, parabola, and hyperbola curves. Experimental results show that among the nonlinear Radon moments, the descriptor extracted by parabola Radon moment is more robust to noise and has more high retrieval efficiency.6A novel and generalized circularity measure is proposed using3D polar-radius-moment and then applied to3D models retrieval due to3D circularity measure problem. The new circularity measure is defined and proved. The circularity values also can be controlled in a fixed range by a suitable value of order according to degree of accuracy or other requirements in actual application. Experimental results show that the new method can discriminate differences between similar shapes. And the performance of proposed method is simple and more efficient in3D model retrieval.7Due to the inconvenience of fast computation of V-system moment, a W-system moment is defined and the corresponding fast computation method is proposed. Combined VDS invariant, the new features are applied to three dimensionality models retrieval. Experimental results show that the new approach is more efficient than orther comparative approachs in describing3D shapes.