Adaptive Multi-Hash Similarity Fusion for Remote Sensing Image Retrieval

With the exponential growth of remote sensing (RS) images, content-based remote sensing image retrieval (CBRSIR) has become an effective method to retrieve target images from RS data. However, existing CBRSIR methods often rely on a single feature, which fails to fully capture the rich and complex visual content. To address this limitation, we propose an adaptive multi-hash similarity fusion method for RSIR. Firstly, deep features and manual features are extracted to provide varying perspectives on the image. Secondly, in order to enhance retrieval efficiency, these features are transformed into low-dimensional hash codes through the probability ordinal-preserving semantic hashing method. Finally, various hash similarities are fused to capitalize on the complementarity of different views and narrow the semantic gap. Given that different hash similarities contribute disparately to the final retrieval result during fusion, determining the weight distribution is critical for enhancing the fusion effect. Therefore, the mean average precision results are utilized to adaptively determine fusion weights. Experiments on two benchmark datasets demonstrate that the proposed method outperforms other RSIR methods. Comparative experiments with exhaustive weighting reveal the simplicity and effectiveness of our adaptive weight allocation method.