Relevance Preserving Projection and Ranking for Web Image Search Reranking With Hierarchical Topic Awareness

Image Search re-ranking is as an effective way to improve the retrieval precision. Image search re-ranking has recently been proposed to refine image search results obtained from text-based image search engines. Most of the traditional re-ranking methods cannot capture both relevance and diversity of the search results at the same time. Or they ignore the hierarchical topic structure of search result. Each topic is treated equally and for certain queries are naturally in hierarchical organization, rather than simple parallel relation. Image Search Reranking (ISR) technique aims at refining text-based search results by mining images visual content. Feature extraction and ranking function design are two key steps in ISR. Here a new re-ranking method Topic-Aware Re-ranking (TARerank) is proposed. TARerank describes the hierarchical topic structure of search results in one model, and seamlessly captures both relevance and diversity of the image search results simultaneously. By set of carefully designed features, through a structured learning framework, relevance and diversified images are modeled in TARerank, and then the model is learned from human- labeled training sample. The learned model is expected to predict re-ranking results with high relevance and diversity for testing queries. We collect an image search dataset and conduct the comparison experiments on it to verify the effectiveness of the proposed method. The experimental results demonstrate that the proposed TARerank outperforms the existing relevance-based and diversified re-ranking methods. Extensive experimental results on real-world datasets show that the proposed algorithms are effective. Moreover, the fact that only relevant images are required to be labeled makes it has a strong practical significance.