Improvement of image retrieval accuracy based on incremental vocabulary tree using chain feature selection algorithm

Abstract

Any task of robot consists of three steps such as recognition, decision and action sequentially. Among them, decision step should be performed after modifying input data for current task as suitable as possible. In the case of data categorization task, we can define classifier which means the decision step module which can determine the class of given query data. Especially, robot system requires correct decision technique for object such as human, material and other environments when some input image informations are given. This is why object recognition field experiences pretty much spotlight recently. Because of fast improvement for information process and data mining skill, main concerning is concentrated on real-time classifier maintaining proper object recognition performance. The one of representatives in real-time classifier is Incremental Vocabulary Tree. There is two characteristics at this structure. First, this classifier doesn`t require pre-defined classes for constructing and training process because it uses hierarchical k-means clustering. So, this classifier is based on contents based image retrieval. Second, when new image is inserted continuously, this structure could increase or decrease the size of node to adapt the information from new query image and maintain own size as possible. But, maintaining proper retrieval performance also requires large number of nodes and that number could change slightly when we perform decision process. This is why incremental vocabulary tree based on local feature have had a problem with the respect of combination global features such as color feature, texture feature and etc. In this thesis, we propose feature selection method for incremental vocabulary tree to select super nodes which can make big difference between training images and fix the size of those useful nodes. Although existing feature selection method have been known, those method could not consider about large number and large size of feature...