Deep rotation-invariant feature map hash for remote sensing image retrieval
Hash generates compact hash code for data representation,and is widely used for large-scale remote sensing image retrieval due to its efficiency.Affected by satellite observations,the same object may appear at multiple angles in different remote sensing images which leads to decline of retrieval performance.To address this issue,this paper proposes deep rotation-invariant feature map hash,i.e.,DRIFMH,including feature map extraction and hash quantization modules.The feature extraction module rotates the feature map at different angles,proposes a feature consistency loss,and maintains the consistency of image features at different rotation angles,overcoming the adverse effects of rotation.The hash quantization module performs binary quantization on image features to generate hash codes,and introduces classification cross entropy loss to enhance the discriminative ability of the hash codes.The proposed method is evaluated on AID and UCMD datasets and compared with multiple hash methods in experiments,and the empirical results demonstrate that the proposed DRIFMH can generate rotation-invariant remote sensing image feature and improve the performance of large-scale remote sensing image retrieval.
国家科技期刊平台
NSTL
万方数据
