Research On Scene Classification Of LDA Based On Visual Dictionary Capacity Automatic Obtaining

It produces large amounts of digital images every day with the rapid popularization and applications of computer technology and Internet techno logy.Faced with huge amounts of image data sets, how to make the computer in accordance with the manner of human cognition to classify the image data sets efficiently has become a hot issue in the study field of image understanding.The classification methods of scene images are diverse, among them, using the LDA model for scene classification has become a research focus at present.Classical method of scene classification based on LDA model is:Firstly, extract SIFT features of all scene images,and K-means clustering algorithm is used to cluster SIFT features of scene images, and structure a visual dictionary, and then calculate Euclidean distance of the scene image SIFT features and visual words in the dictionary, to generate the word frequency matrix. Finally, using LDA model to learn and complete the scene classification. Taking the K-means clustering algorithm to obtain the visual dictionary capacity, it needs factitious trial repeatedly to determine the reasonable visual dictionary capacity,so efficiency is very low. In allusion to this problem, this paper puts forward to using the AP clustering algorithm automatically obtain reasonable capacity of visual dictionary, and then implements the scene classification of LDA model, improves efficiency of scene classification.In this paper, the main work is as follows:Firstly, this paper introduces the background and research significance of the scene classification,analyzes the research status of scene classification, and introduces research work and main research results in this paper.Secondly, this paper introduces the related theory of scene classification method.It gives the whole framework of scene classification,and states the detailed process of SIFT features of scene images extraction;It lists the types of clustering algorithm,the K-means clustering algorithm and the AP clustering algorithm are introduced in detail, and analyzes the advantages and disadvantages of these two kinds of clustering algorithms.Thirdly, this paper proposes the scene classification method of LDA based on visual dictionary capacity obtaining automatically.We use the SIFT algorithm to extract SIFT features of images, then it obtains the capacity of visual dictionary with the K-means clustering algorithm and AP clustering algorithm respectively, and builds the visual dictionary to generate the frequency matrix from images SIFT features and the words in the dictionary.Finally,the LDA model was used to learn the potential topic distribution to implement the scene classification.The experimental results show that the scene classification method we proposed is more efficient.Fourthly,this paper analyses the results of the experiment.It compared the experimental results between the scene classification method of LDA based on K-means clustering algorithm and the scene classification method of LDA based on AP clustering algorithm, and find that scene classification method of LDA based on AP clustering algorithm can obtain reasonable visual dictionary capacity more easily, and the accuracy rate of scene classification can be over 79%. While using the scene classification method of LDA based on K-means clustering algorithm, there is no predictable pattern in relation curve between classification accuracy rate and visual dictionary capacity. And the highest precision of the classification can only be 78.10%. Then, the influence of topic number of LDA model to scene classification performance,the experimental results show that the classification accuracy is higher when the topic numbers are relatively smaller;the two hyper-parameters α and β of LDA model have no influence in classification efficiency; the variation of classification accuracy is irregular with the hyper-parameter a changing,but increasing the hyper-parameter β decreased classification accuracy. Finally, by the analysis of confusion matrix of image scene classification results, the research results of this article state that the classification accuracy rate of indoor scene is quite low with LDA model.Finally, the research results of this article make a summary, points out the main contribution in this paper, and gives the recommendations for further research..