Orthogonal Image Moments And Its Application In The Accurate Positioning And Product Inspection

Machine vision technology is the emerging technology based on image processing and pattern recognition. The target image can be obtained through the image capturing device in a typical machine vision system, afterward the information is passed to the image processing section, then several functions such as Pattern recognition, online Inspection and target tracking can be achieved, the implementing agencies can be driven to complete the relevant action. It has the advantages of non-contact, high precision and rapidity, and is the important component of advanced manufacturing.Image processing technology which has a direct impact on the machine vision system is research hotspot in the machine vision, and the image moments are the expression of the image shape feature, which describes the characteristics of the image area. Orthogonal image moments are an important branch of image moments, compared with non-orthogonal image moments, which have a number of significant advantages.The calculation of traditional orthogonal continuous image moment needs the remapping of the coordinate space and the treatment of integral approximation. On the basis of previous research, the method of accurate calculation of the pseudo-Zernike moments is proposed. Firstly, pseudo-Zernike moments are converted into a linear combination of Fourier-Mellin moments, then the combination of a number of rectangles, triangles and fan-shaped region is used to express the entire computational domain, afterward, in order to reduce the computational complexity a recursive expression is deduced through the integral relationship of the trigonometric functions to solve the above-mentioned shape of the region's moment integral. In addition, the accureate computation of Pseudo Zernike moments based on the recursion is introduced, which uses the recursive relations to calculate the moments for the sake of reducing computation time.The calculation of traditional discrete orthogonal image moments don't require the treatment of discretization and integral approximation, and is usually solved by iteration and symmetry relations. However, the increase of the number of iterations will result in dramatic increase of transmission error in solving large-scale image of the high order discrete orthogonal image moments. Considering Krawtchouk moments as the object of study, the symmetry relations of Krawtchouk Polynomials in different parameter P are analyzed, and a novel bi-recursive algorithm is proposed to calculate Krawtchouk polynomial in which the maximum number of iterations will be reduced to the original number of general. The new algorithm can improve the accuracy of the calculation of the Krawtchouk moment. In addition, the calculation of subparagraph can be taken account into this algorithm, which further reduces the number of iterations and the impact of the transmission error. The radial orthogonal image moments which have the characteristics of the radial moments are a special class of orthogonal image moments, and the amplitude of these moments has nothing to do with the angle of rotation, and therefore the rotation invariant can be directly obtained. The calculation of radial continuous orthogonal image moments requires the treatment of discretization, which causes the discretization error. The radial Bi-discrete Fourier transform is presented on the basis of the derivation formula of the radial continuous orthogonal moments. First the mapping relationship of the Cartesian coordinates and polar coordinates is studied, and then a polar coordinate system with orthogonal discrete Fourier expression is built, finally the orthogonal cosine transform is used to replace the Fourier transform in order to avoid complex operations and improve computational efficiency.The choice of image feature is the key factors in target recognition. Zernike moments as the radial orthogonal continuous moments have a direct response of the regional characteristics of the target, and are suitable for the description of the complex boundary goal. The cutting radius of Zernike moments directly affects the stability of the invariants, and the method of calculating a cutting radius is proposed. Because BP neural networks is easy to fall into local extreme value and is sensitive to initial value, particle swarm algorithm is used to optimize the weights and thresholds of neural networks, and chaos operator is utilized to initialize weights and thresholds of neural network, then particle swarm takes advantage of the chaos mechanism avoiding precocity.The main task of image processing of dynamic accurate location systems is to strike the image of the target location. The target characteristics generally include edge features and point features, in which edge features need a larger handling capacity compared with point features. A point feature-based matching method is introduced. Firstly, adaptive median filter is used to deal with salt and pepper noise, and then two-level Zernike moments is utilized for the edge of the sub-pixel detection, followed by the modified method of the curvature scale space is applied to extract corner points, and finally the positioning of the target is completed through the local and global match.Character and defect are the common projects of product inspection. Zernike moments and skeleton features are chosen as features in the character recognition, and a two-tier cascade identification method is used to ensure the recognition performance and identify efficiency. Firstly, the mean gray level grayscale vector is used to complete the rough match, and finally orthogonal Fourier-Mellin moments are utilized for fine matching. This new method can take into account the accuracy and rapidity of dectection.