A Feature Vector Approach for Inter-Query Learning for Content-Based Image Retrieval

Abstract

Use of relevance feedback (RF) in the feature vector model has been one of the most widely used approaches to fine tuning queries for content-based image retrieval (CBIR). We propose a framework that extends RF to capturing the inter-query relationship between current and previous queries. Using the feature vector model, this avoids the need to "memorize" actual retrieval relationships between actual image indexes and the previous queries. This approach is suited to image database applications in which images are frequently added and removed. In the previous work, we developed a feature vector framework for inter-query learning using statistical discriminant analysis. One weakness of the previous framework is that the criteria for exploring and merging with an existing visual group are based on two constant thresholds, which are selected through trial and error. Another weakness is that it is not suited to mutually interrelated data clusters. Instead of using constant values, we have further extended the framework using positive feedback sample size as a factor for determining thresholds. Experiments demonstrated that our proposed framework outperforms the previous framework.