摘要
哈希技术采用紧致哈希码表示数据,因其高效性被广泛应用于大规模遥感图像检索任务。受卫星观测影响,同一地物在不同遥感图像中呈现不同角度,导致检索性能下降。为解决该问题,该文提出深度特征图旋转不变哈希方法(DRIFMH),包括特征提取、哈希量化2个模块。特征提取模块对特征图进行不同角度旋转,提出特征一致性损失,使不同旋转角度的图像特征保持一致,克服旋转带来的不利影响。哈希量化模块对图像特征进行二值量化,生成哈希码,引入分类交叉熵损失,提升哈希码的鉴别能力。该文选取经典遥感图像数据集AID、UCMD作为实验数据集,将DRIFMH与多个哈希方法进行实验对比,结果表明DRIFMH能够生成旋转不变的遥感图像特征,提升大规模遥感图像检索性能。
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.
作者
胡明浩
张博文
沈肖波
孙权森
Hu Minghao;Zhang Bowen;Shen Xiaobo;Sun Quansen(School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China)
出处
《南京理工大学学报》
CAS
CSCD
北大核心
2024年第4期434-441,共8页
Journal of Nanjing University of Science and Technology
基金
国家自然科学基金(62176126)
江苏省自然科学基金(BK20230095)
中央高校基本科研业务费专项资金(30921011210)。
关键词
遥感图像检索
哈希
特征图
旋转不变性
remote sensing image retrieval
hash
feature map
rotation-invariance
