A hierarchal BoW for image retrieval by enhancing feature salience

Retrieving images with multiple features is an active research topic on boosting the performance of existing content-based image retrieval methods. The promising bags-of-words (BoW) models involve multiple features by applying feature fusion strategies in the early stage of image indexing. However, due to the different data forms of features, a simple joint may not guarantee a high retrieval performance. Moreover, a fused feature is not flexible enough to adapt to the variety of images. In order to avoid the submergence of feature salience, this letter proposes a hierarchal BoW to represent each feature in an individual codebook for obtaining the undisturbed ranks from each feature. Moreover, for feature salience enhancement, a query model based on ordinary-least-squared (OLS) regression is established for rank aggregation. The query model weighs each feature according to its retrieval performance and then selects the target images. The experimental results demonstrate that the proposed method improves the accuracy compared to the state-of-the-arts, meanwhile it maintains the stability.