The Improvement Of Bag-of-visual-words Model And Its Application Research In Images Classification

In recent years, with the Internet continuing to surge in digital images, it has clearly becomeone of the problems that need urgent solutions how to classify the massive images in hugedatabase quickly and accurately. Many existing images classification methods generally obtainthe visual features information by extracting the low-level images features, but there is differencebetween the different images, the numbers of low-level features extracted from different imagesare generally unequal. To improve the performance of comparison between different images, theresearchers apply the bag-of-words model in the field of text classification to the imagesclassification field, known as the bag-of-visual-word model. In bag-of-visual-words model, thelow-level images features are extracted and descripted, the extracted features are quantified toget a visual dictionary, then each image is expressed as a frequency histogram based on the samedimension visual words, the vector is a description of the image. Finally, the description vectorsof images are substituted into the classifier for classification.This paper focuses on the bag-of-visual-words model for the images classification problem,studying the features extraction methods, the dictionary generation methods and the imagesrepresentation methods of the bag-of-visual-words model, the specific research results are asfollows:(1)To avoid the blindness of selecting the gradient direction discrete precision in routinelybuilding bag-of-visual-words model, as well as low classification precision results from poorbag-of-visual-words model, the purpose of this paper is to study whether there is an obviousthreshold of gradient direction discrete precision during extracting the image s features inbag-of-visual-words model, in order to obtain a threshold which can lead to the bestbag-of-visual-words model. Based on fast local descriptor oriented dense feature extraction, alarge number of experiments are done with different gradient direction discrete precisionrespectively. Experimental results fully show that24is the clear and unified threshold, thebag-of-visual-words model which choose the threshold can get the highest correct classificationresults.(2)The Scale-Invariant Feature Transform algorithm (SIFT) for bag-of-visual-words modelmay fail to extract high discriminating and representative features. To solve this problem, a newalgorithm based on corner feature weighting for bag-of-visual-words model images classificationmethod was proposed. The fast dense mesh generation method was used to extract grid SIFT, inorder to obtain more significant and representative points, the Harris corner detection algorithmwas adopted to detect the images corner, then the corner SIFT features were extracted in the lightof each position, meanwhile，a certain weight for corner SIFT feature was set on the basis ofdegree of corner point. This can make the corner SIFT features more prominent compared withthe grid SIFT features, so that the images can achieve high-quality description. The experimentalresults show that the bag-of-visual-words model constructed by the proposed approach workswell for images classification. (3)For lack of spatial pyramid bag-of-visual-words model to express the semanticdistribution relationships between the local features, we propose a new spatial pyramid bag-of-visual-words model method for images classification based on semantic phrases. Firstly, thelocal features are mapped to visual words with certain semantic information semantic phrases areconstructed by making statistics of the distribution of other relevant neighborhood characteristicswithin the range of local features. Secondly, sparse coding is adopted to train a semanticdictionary by semantic phrases, the images are represented as spatial pyramid sparse histogramvectors based semantic dictionary. Finally, the vectors represented images are substituted into theclassifier for training and testing. The experimental results show that this method can greatlyimprove the accuracy of the images classification.