Feature Fusion Based on the Average Precision in Image Retrieval

In content-based image retrieval, image has various inherent aspects which reflect its contents, therefore how to organize and utilize these contents effectively to improve the retrieval performance is a valuable research topic. In this paper, a method of measuring the complementarities between two feature spaces is proposed, based on which the fusion feature set can be selected effectually, and the experimental results are positive. At the same time, a linear fusion method based on the average precision of features is proposed. Extensive comparisons against several methods, such as flat model and rank-based linear fusion are performed. Experiments are carried out on a large-size heterogeneous image collection consisting of 12000 images and the results demonstrate the effectiveness of the proposed method.