Research On Image Categorization Based On Bag-of-words Model

With the development and popularization of the network and the maturity of the cell phone camera technology,digital image is widely used in people's daily life.The fast and accurate classification of the huge amount of image data is an important subject of image processing field and the bag of words model which is based on semantics is an important direction of the research of the digital image classification.It draws extensive attention and is in widely used for its achievement of favorable effect in image classification based on content.Its application is very extensive in the field of information security,including how to identify the sensitive,illegal images and prevent teenagers on bad information browsing,etc.After deep and systematic research in image classification it is discovered that the traditional method which first extracts images by SIFT feature,then makes image dictionary by using K-means clustering,further more comes to bag of words by applying histogram and at last utilizes unsupervised classification is not accurate enough when dealing with tremendous amount of images.Three aspects of work are done in this dissertation in order to improve the accuracy of image classification.(1)Bag of words is a high-dimension feature vector.Its image classification results depend highly on the performance of the classifier,with different classifiers being applicable for different feature vectors.On the basis of the bag-of-visual-words,this paper carries out the research and experiment on KNN classification algorithm,deep learning neural network and support vector machine SVM.Meanwhile,it makes in-depth analysis on the classification of the high-dimension vector of the bag of words.Few samples are enough for confirming that the SVM of classification hyperplane is superior to KNN classification algorithm and deep learning neural network in terms of the classification result.(2)SIFT feature extraction is to seek for the extreme point in the local space of the image,having the feature of zoom and rotation invariance.It could adapt to the changes of perspective,light and color.It is the description of local characteristics.The performance of its feature extraction result directly affects the separating capacity of the bag of words.This paper uses the overlapping SIFT feature to extract the image information,extracts more image features and extracts the information in the bottom layer of the image more comprehensively.In this way the developed bag of words could represent the image semantics between and achieve better result.(3)As the visual vocabulary,the bag of words is important for the image classification.To further improve the image classification result,it brings the histogram similarity measurement into the bag of words to develop the histogram intersection kernel.On the basis of the bag of words and the calculation of its intersection kernel,it improves the distinction degree of the bag of words generated by the image of different types.Besides,it uses the SVM classification method and makes experiment.The experiment demonstrates that the semantic expression of bag of words of histogram intersection kernel on the image is superior to the ordinary bag of words,which further improves the classification result of the image.The detailed and thorough experiments are done on Caltech256 standard image classification data.The experiment results show that the proposed three improved algorithms in this dissertation can effectively improve the accuracy of image classification.