A Novel Retrieval Refinement and Interaction Pattern by Exploring Result Correlations for Image Retrieval

摘要

Efficient retrieval of image database that contains multiple predefined categories (e.g. medical imaging databases, museum painting collections) poses significant challenges and commercial prospects. By exploring category correlations of retrieval results in such scenario, this paper presents a novel retrieval refinement and feedback framework. It provides users a novel perceptual-similar interaction pattern for topic-based image retrieval. Firstly, we adopts Pairwise-Coupling SVM (PWC-SVM) to classify retrieval results into predefined image categories, and reorganizes them into category-based browsing topics. Secondly, in feedback interaction, category operation is supported to capture users' retrieval purpose fast and efficiently, which differs from traditional relevance feedback patterns that need elaborate image labeling. Especially, an Asymmetry Bagging SVM (ABSVM) network is adopted to precisely capture users' retrieval purpose. And user interactions are accumulated to reinforce our inspections of image database. As demonstrated in experiments, remarkable feedback simplifications are achieved comparing to traditional interaction patterns based on image labeling. And excellent feedback efficiency enhancements are gained comparing to traditional SVM-based feedback learning methods.