Researches On Kernel Machine Learning Methods And Its Application In Vision Inspection

During the printing process, the ink picks, color mixing, pulling streaks, color run,missing print factors lead to the non-conforming products. One of the most effectiveways to improve productivity is the detection of product quality using vision technologyand automatic removal of non-conforming products. Given light source, cameras andother hardware conditions, the performance of visual detection technology depends onthe image processing system. The processing method based on kernel machine learningis the fusion of statistical learning theory and kernel technology, which is effective indealing with non-linear data. Under the support of the project "Algorithm research ofhigh precision printed quality inspection", based on in-depth analysis of kernel machinelearning methods, combined with the application background of high precision printedmatters visual inspection, in order to improve the detection performance, focus on theimage restoration algorithm, the training set selection and the learning of kernelfunction and parameters.The main contents can be summarized as follows:1) Basis for the further research on kernel principal component analysis, aiming atthe fuzzy problems of restoration image in quality detection of printed matters,proposed a new image restoration algorithm. The algorithm was designed under thehypothesis that the image samples’ distribution satisfied the local linear properties. Andusing the image samples’ linear distribution characteristics of the feature space, thenearest neighbors were determined. On the linear geometry consists of nearestneighbors, took the residual minimization as optimization objective, and using the fixedpoint iterative algorithm to calculate the optimization. The convergence analysis of thealgorithm was provided. The numerical experiments were carried out to validate therationality of the algorithm. The experimental results showed that reconstruction errorreduced by half compared with the existed algorithm.2) In order to improve the calculation speed of image processing, a samplingalgorithm based on the spatial distribution of the image samples was proposed. Thesampling algorithm not only considered the density distribution in the database, whilethe boundary image samples also were selected. When the sample selection algorithmwas used in image processing system, under the signal to noise ratio approximately the same, with the training sample number from500being reduced to100, training timewas shortened by99.2%, detection time was shortened by96%.3) In order to improve the calculation precision of image processing, the samplingalgorithm based on kernel alignment was designed. The sampling algorithm tooksample’s contribution to the kernel matrix as the choosing standard. The most significantimage sample with the greatest contribution to the kernel matrix was selected one byone from the large data set. Selected significant samples constituted the training set. Thesampling algorithm is convenient for new sample selection and addition, so thesampling process can be completed online. When the sampling algorithm was used inimage restoration, under the same amount of computation, the average reconstructionerror was reduced from1312.7to462.9.4) An algorithm for learning the kernel function and its parameters was designed,which took the standard inner product matrix as learning target. The algorithm can alsobe used for multiple regression and unsupervised problems. Kernel alignment caneffectively measure the similarity between kernel function and target function. Based onthe property, a learning algorithm was proposed for maximization of kernel alignment.The theoretical analysis and research showed that kernel alignment functions containedthe combination coefficients and Gauss bandwidth two variables have at most twostagnation points. An iteration algorithm was designed for calculating the stagnationpoints. Two experimental results showed that the reconstruction results using thelearned kernel function and parameters were closed to the statistically optimal.5) For the kernel machine learning, three extension directions learning vectorvalued kernel function, learning semi-definite positive matrix function and multiplekernel learning were in-depth study. The method of learning matrix valued function wasproposed. Experimental results showed that the learned vector valued function andmultiple kernel function were effective for solving the multi-class problems, de-noisingof multivariate data and multiple regression problems.