A new analysis framework for relevance feedback-driven similarity measure refinement in content-based image retrieval

Many recent content-based image retrieval techniques utilize relevance feedback (RF) from the user to adjust the system response to better meet user expectations. One school of RF-based methods uses a weighted Minkowski distance metric to assess similarity, and adjusts the weights to refine query response. A new method of estimating these weight vectors is presented which outperforms existing methods, particularly for the important case of limited training data. A new objective function is presented for an iterative optimization routine which more closely aligns optimization goals with true system goals. A new analysis framework is presented in the derivation of this technique which is useful for understanding the limitations of many RF methods.