Research On Image Classification Based On LDA And Active Learning

The development of computer and Internet technology has let us enter into a world based on image building. Huge amounts of image information make people’s life more rich and colorful, however, diversification and confusion of information let people can’t get what they need in a timely and effective manner. Owing to this, classifying image effectively, to make originally messy information concise and consistent, becomes the pressing needs of the masses. As a bridge of communication between image and human cognition, image classification technology provides an effective way to solve this problem.Firstly, some basic methods of image classification are analyzed in this thesis, and then based on the SIFT features of images, the subject about probability topic model, active learning and sparse representation for image classification are discussed.Classification based on probability topic model, looks image as document, creates visual vocabulary with the thought of clustering SIFT feature structures, and then determines Dirichlet distribution model’s topic number restrict to the category of the images to be classified, obtains topic distribution vector of all images by Gibbs sampling, with the help of probability maximization criterion, completes image annotation and classification. During the time of clustering, k-means and locality sensitive hashing(LSH) applied to large capacity high-dimensional data have been compared. The results show that LSH can improve method’s efficiency greatly.The task of image classification based on active learning is to solve the problem of costs existing in large capacity sample’s tagging. Based on Gaussian process, starting from the presentation theorem, it discusses two kinds of sampling methods suitable for sample query. The two methods, one kind considering the weight change of the model after example sampling, is called gp-weight; the other one taking change of the overall model into account, is called gp-impact. Upon the example selection, and then trains a binary support vector machine classifier with the new sample set. The results show that gp-weight and gp-impact have strong competitiveness compared to some general methods.Image classification based on sparse representation references the basic idea of neurons’ highly sparse respond to the specific stimuli outside world. This article begins with the SIFT feature extraction of images, on the basis of an over-complete dictionary, completes sparse representation on image with norm optimization method, reconstruct original images, and then use residual minimization criterion to determine each image’s category. In addition, it proposes two kinds of dictionary, one is composed of training samples, the other is completed by online learning, these dictionaries’ effect on subsequent computation of signal’ sparse vectors are verified by experiments.