Research On Boxing Object Location Algorithm For Bag Of Visual Word Model

This paper focuses on the approach with respect to the image classification task, ofwhich theobjective is to search for the similar images in the large image collections. As asignificantfundamental research in computer vision, dozens of related algorithms andharvestable resultshave been reported in the past decade.In the image classification task, the images are firstly represented by the discretefeature vectors. Reference to the Bag of Words (BoW) representation used in textualinformation retrieval, the state-of-art Bag of Visual Words(BoVW), which describes theimage as an orderless collection of thelocal features, has been widely studied. Most of therelated BoF-based researches focus onlearning more discriminative visual vocabulary. Fromanother perspective, a methodology thatjoints the localization and the classification task hasbeen paid more attentions in recent years,following the intuition that knowing the objectlocation can be helpful for the classification. Therefore, the capital challenge of this kindapproach becomes to box the object location in the image firstly.In this paper，we suggest an Object-Shrunken (OS) algorithm to handle the imageclassification task. Unlike the prior art，this paper considers the foreground or the objectlocation in the image for more discriminative image-level representation. TheObject-Shrunken algorithm suggests a straightforward procedure to box the object location.It first proposes a Weighted Local Outlier Factor (WLOF) to remove all the interest pointoutliers， and then positions the object location in terms of the distribution of the restinterest points. We evaluate the proposed algorithm on the well-known dataset Caltech-101.The resulting Object-Shrunken algorithm outperforms the state-of-art approaches in theimage classification task.Currently, the most of these joint approaches need sufficient supplies of the objectclasses tolearn various object models, and further locate the object location. In this paper,we suggesta straightforward method to box the object location, which is so calledObject-Shrunken (OS)algorithm. For the final goal of classification, the Object-Shrunkenalgorithm does not focus on extremelyaccurate localization, but shrinks to the rough objectbounding box for more discriminative image-level representation. It assumes that the interest points detected in the image are almost occurredaround the object. Thus, we canposition the object location after removing all the interestpoint outliers, which areconsidered as noise. Although the Object-Shrunken algorithm is straightforward, itiseffective to improve the classification performance. A series of numerical experiments ontheCaltech-101dataset show that the proposed Object-Shrunken algorithm outperforms thestate-of-art approachesfor the image classification task. In summary, the main contributionsof this paper are as follows:1.We suggest a straightforward method to box the object location in order to improvetheclassification performance, which depends on the interest point distribution.2.We weight the Local Outlier Factor (LOF) to judge whether an interest point is anoutlier inorder to obtain more shrunken object bounding box.