The Discovery of User Concept Region Based on Multiple Instance Learning
|School||Harbin Institute of Technology|
|Course||Computer Science and Technology|
|Keywords||Content-based Image Retrieval Multiple Instance Learning User Concept Region High-level Semantic Concept|
How to learn the user semantic concept in content-based image retrieval is an effective solution for“semantic gap”problem. Region of interest is the focus of the users’attention in an image and is better to represent the user concept than other regions. Learning the user semantic concept based on user concept region is the main research content of our paper.Multiple-instance learning (MIL) is a new machine learning mechanism that is different from the original machine learning mechanism. In the framework of MIL, it tries to learn a function based on the unlabeled instances produced by labeled training samples to predict the unknown samples effectively. In CBIR, images are segmented into regions to form the instance space in which the“Ideal instance”is searched based on the Diversity Density algorithm, and the“Ideal instance”stands for the user concept region. User concept region can be effectively learned or discovered based on MIL.The main contribution of the paper is offering a solution to the discovery of the user concept region based on MIL, and constructs the content-based image retrieval system to verify the efficiency of the proposed algorithm. The main research work is list as follows:Firstly,from the view of machine learning, we research into the machine learning mechanism. And we focus our attention to the theory and applications of MIL, especially in CBIR.Secondly, from the tradeoff the targets and noise, we propose the algorithm for multi-scale instance production based on image partition utilizing sliding window.Thirdly, from the application view point, we construct the content-based image retrieval system based on MIL framework. In order to evaluate the system, we do the experiments based on the Corel 10000 image database. We adopt the P@N to evaluate the precision of the system and get the results that the precisions reach to 56.6%、42.6%、36.6% in the cases of top10、20、30 images that is superior to QVE relevance feedback method while consuming the less time complexity Fourthly, the concept region discovery based on MIL makes better performance in precision than the CKNN-ROI method and Saliency Map method.