Large Scale Object Classification And Retrieval Based On Bag Of Visual Words

Telephones with a camera and digital computers rapidly increase, which makes the explosive increase of the number of digital images. Thus, the research about technologies of dealing with large scale images becomes extremely active. Especially, large scale object classification and object retrieval are two critical directions.Researchers introduce the idea of Bag of Words(BOW) into the field of images, and develop the Bag of Visual Words(BOVW) model. BOVW model has two stages: the Stage of Generating Vocabulary and the Stage of Quantifying Local Descriptors. The stage of generating the vocabulary selects a part of local features by training the library of local features, which serve as the words of vocabulary. For the problem of content based object classification and retrieval, the vocabulary generated by methods based on clustering is more representative and discriminative.Therefore, the method of quantifying local descriptors in the stage of quantifying local features has a significant impact on the performance of the BOVW model in object classification and object retrieval. For the high dimensional, how to efficiently save and match the high dimensional BOVW features make challenges to the large scale image retrieval.For the problems and challenges above, the main work of this paper is as follows:Firstly, we study various methods of quantifying in the bag of visual words model and the experimental results show that radius-based bag of visual words model is inferior to bag of visual words model of KNN quantifying. Thus, we propose the rate of local feature usage and explain the rationality of the experimental results.Secondly, utilizing the sparse property of high dimensional BOVW and the efficiency of the storage of inverted file index and querying it, we present a new Compressed Bag of Visual Wordsmodel(CBOVW). CBOVW greatly decrease the dimension of BOVW, which reduces the computational complex in the stage of matching.Thirdly, to make use of the efficiency of hashing based methods for large scale object retrieval, we aggregate spectral hashing(SH) into the process of matching CBOVW, and propose a new algorithm named a fast compressed bag of visual words(FCBOVW).Our experiments are on the object image databases PASCAL VOC 2006 and University of Kentucky Recognition Benchmark. We test the performance of RBBOW in object classification and the capability of CBOVW and FCBOVW in object retrieval on the two databases, respectively.The experimental results show that the speed of CBOVW and FCBOVW in object retrieval is competitive with other methods, and the method of effectively computing eigenvector should be considered to improve the accuracy of object retrieval.