Query-Adaptive Feature Fusion Base on Convolutional Neural Networks for Remote Sensing Image Retrieval

Abstract

Content-based Remote sensing image retrieval (CBRSIR) becomes important research with the volume of remote sensing images rapidly expanding. Many image features have been proposed for CBRSIR, hence it has become a big challenge to effectively fuse these features for alleviating the huge variation in retrieval performance among different image queries when a single image feature is used. We proposed a query-adaptive feature fusion method based on a convolutional neural networks (CNN) regression model. We use the CNN regression model to estimate the DCG value for each feature and assign different features with different weights for each query according to these DCG values. Meanwhile, we use the image-to-query-class distance to further improve retrieval performance. Experiments on UCMD show that the proposed method can improve the CBRSIR performance.