Visual Dictionary Pruning Using Mutual Information and Information Gain

Abstract

AbstractFeature selection methods are often applied to many machine learning problems, one of the applications involves selecting most informative Visual Words for image categorization task. In Bag of Visual Words framework, image is represented as vector of frequencies of Visual Words, typically of length from hundreds to thousands elements. A dictionary of Visual Words is produced from image keypoints detected by SIFT algorithm and quantized into words by k-means clustering. In the paper we use Mutual Information and Information Gain as methods for selecting these words that are the most important for efficient image classification. There are four novel methods, which expand use of classic Mutual Information and Information Gain in line with our previous feature selection methods. We consider two basic selection strategies: one-vs-all and one-vs-one, as well as multi class and multi attribute value problems. The experimental session we have conducted has shown a positive effect of our modification, when applied to image classification by Support Vector Machines. The results showed that visual word selection based on modified Mutual Information in most cases wins over methods based on Information Gain.KeywordsVisual Bag of WordsFeature SelectionSupport Vector MachineMutual InformationInformation Gain