Multiple Kernel Learning For Image Classification

In recent years,with the rapid development of computer visual technology,image classification has been cared by researchers as a classic problem.At the same time,in the respect of image characteristics,the researchers put forward various kinds of visual features,which can more accurately describe images from different aspects and angles.In the traditional image classification methods,we usually use a feature or experimental ways to distribute the weight of each feature.If you want to make better use of the feature to enhance the image description ability,it requests us to combine them in a suitable way.Multiple nuclear learning is an extension based on single kernel learning.Multiple kernel learning can obtain the optimal combinations of kernel in the current task according to different learning tasks.In this paper,we use the late features' fusion with the form of multiple kernel learning to combine various characteristics of the effective use,using the complementary information to enhance image classification ability.This paper is mainly for researching the image classification problems based on multiple kernel learning,and solving the problems of multiple kernel learning in which study time cost is high and can't get a sparse solution when applied to image classification.At the same time,based on the features that image representation has rich forms and big differences in the characteristics,combination with clustering methods further improve the accuracy of image classification.In this paper,the main work is as follows:(1)Proposing the Multi Kernel Optimization Strategy based on Kernel Similarity.Considering multiple kernel learning has a high learning complexity,through analyzing the relationship of multiple kernel learning and Ensemble Learning,and referring from the Ensemble Learning strategy,we choose parts of high-quality kernel in the initial stage of multiple kernel learning.In order to get accurate assessment criteria,we explore various types of evaluation method based on the kernel similarity.Finally,with the form of clustering,we combine the similarity between kernel and the classification of nuclear to choose high-quality kernel.Using the alternative kernel after optimization for multiple kernel learning,we can acquire smaller time complexity.(2)Proposing the Multi Kernel Learning Based On Kernel Weight Adaptive Adjustment.We combine the clustering and the multiple kernel classification learning,using data obtained by clustering to distribute information for doing multiple kernel learning classification.In the solving stage of kernel weight,we can use the optimization problem based on kernel similarity to accelerate the solving process.On the basis of the test,we can adjust the number of sample classification,obtain details of the implicit structure in the sample set better,and improve the effect of the final image classification.(3)Testing for the two methods in Fifteen Scene Categories image data sets and 12306 image data sets.Experiments show that Multi Kernel Optimization Strategy based on Kernel Similarity can effectively improve the sparse solution and reduce the learning time under the situation of no loss of classification accuracy.While we explored optimizing proportion to the influence on the kernel optimization strategy in the experiments,we verified the Multi Kernel Learning Based On Kernel Weight Adaptive Adjustment can obtain better classification effect and solving speed than other similar algorithms.