Kernel ELM Theory And Algorithms And Application In Image Processing

Data based machine learning studies the patterns of data, implements the data classification and prediction. As a class of machine learning algorithms, extreme learning machine (ELM) is a kind of simple and effective single hidden layer feed-forward neural network(SLFNs) learning algorithms, which has attracted more and more researchers. Traditional neural network learning algorithms(such as BP algorithm) have to tune a large set of network training parameters. Ad-ditionally it is prone to a local optimal solution. On the other hand,only the number of hidden layer nodes of networks requires to be set in extreme learning machine without the need of tuning the weights of the input and hidden layer bias,which leads a unique optimal solution. Kernel ex-treme learning machine brings the kernel function into extreme learning machine and gets the least square optimization solution, which means theoretically the proposed algorithm is more stable and has better generalization performance.In the kernel ELM framework, based on the existing research works, several efficient methods and applications of image processing in this dissertation are summarized as follows:(1) For extreme learning machine, the sparse approximation of extreme learning machine has been put forward. For a massive data set in regression and classification problems, the proposed algorithm will use iterative computation instead of solving kernel matrix inversion, which solves a linear equations of relatively small and completing training of extreme learning machine. It reduces the complexity of solving the nuclear matrix inverse and improves the computational efficiency. A large-scale data samples has simulated in the regression and classification problems. Compared with the other existing method, it shows the effectiveness of the proposed mechanism.(2) Because the single kernel cannot satisfy the application requirements such as heteroge-neous data,irregular data, and uneven samples etc.,extreme learning machine base on multi-kernel learning classifier is proposed. In multiple kernel framework, multiple kernels and the weight coefficients map to the new feature space with feature mapping and the weight coefficients are obtained by optimized kernel learning. The different optimization algorithm QCQP and SILP are compared using the standard data set.From the experimental results, the proposed approach shows many learning advantages.(3) Face gender recognition based on evolutionary extreme learning machine is proposed. The curvelet coefficient is extracted from the original image,which reduces the high dimension space by the incremental bilateral two-dimensional principal component analysis (IB2DPCA). At last, it classifies the gender with the evolutionary extreme learning machine. Using the standard face data set,compared with the traditional face gender recognition method, the proposed algorithm is fast with a higher classification accuracy.(4) The method that realizes image super-resolution restoration research based on sparse cod-ing and kernel extreme learning machine is presented. Image super resolution representation in-cludes two phases:training phase and testing phase.In the training phase, with the kernel extreme learning machine, it is mapped from low resolution image to a high resolution that gets the sparse coefficient of coupled dictionary. In the test process, using sparse coefficient of coupled dictionary, it reconstructs the image and gets the super-resolution image with kernel ELM. The algorithm overcomes the bottleneck problem in the sparse dictionary learning and effectively improves the quality of the restored image.(5) Kernel extreme learning machine as a classifier and application in the surface defects of steel ball is presented. Because the steel ball is very small, smooth surface, and light intensity, it is not easy to be located. Using the traditional artificial detection method, it is very hard to ensure its stability and accuracy. The machine vision technology is used for steel ball surface detection, and kernel classifier is adopted for classification design. Compared with the error BP algorithm, experiment results show that the proposed algorithm improves the accuracy and reliability.