Research On Image Classification And Retrieval Based-on Bag Of Features

Content-based image classification and retrieval plays an important role incomputer vision. The Bag of features approach, which compact the whole local featuresas a group, eases that problem to a great extent. According to visual codebook, localfeatures are quantified and labeled as visual words, and set of local features are mappedinto a distribution histogram of key words. The Bag of features method is the result ofapplying the idea in text semantic processing into image processing, and makes a greatimprovement in the efficiency of image classification and retrieval.In this paper, we first make an introduction to different kinds of image features,then give a review of the current situation and research trend of image semanticunderstanding. This paper deeply studies the construction of BOF, as well as thetechnology of classification and retrieval based on BOF. Moreover, we also studyvarious factors that affect the performance of BOF, including the detector of keypoint,the size of the codebook, the weighted strategy of visual vocabulary, and the kernel ofSVM classification. We give an estimation of these factors in detail. As the BOFrepresentation is always of high dimensionality, this paper also present a new approach,the muli-resolution BOF, to speed up the matching of BOF. The main idea is toconstruct lower resolution vectors according to the original BOF vector. Through thecomparison at lower resolution, we can filter a lot of candidate vectors at higherresolution, as as to improve the searching efficiency of BOF vectors. We apply twodifferent methods to construct the muli-resolution BOF: the uniform quantizationmethod and the non-uniform quantization method. The VA-file method is also used inthis paper to ameliorate the I/O and retrieval efficiency in large scale of image datasets.