有监督相似性保持的深度二阶哈希方法

Supervised Similarity Preserving Deep Second-order Hashing

近年来深度哈希方法因其存储效率高和查询速度快的优势在大规模图像检索领域受到了广泛关注。为改善深度成对有监督哈希方法在图像检索上的性能,从提高深度哈希获取图像特征的全局性和同类样本相似性角度出发,提出了一种有监督相似性保持的深度二阶哈希方法。该方法采用成对样本图像进行特征建模,并利用协方差估计来捕获样本图像的深度二阶信息,以获取具有良好全局表达能力的深度二阶哈希码;在此基础上,借鉴类哈希近似二值化来解决哈希映射过程中的非凸性问题,以更好地避免量化误差,同时基于多损失函数集成思想构建类别监督和相似性保持的联合约束,进而采用交替迭代的优化方式实现网络的端到端训练,最终确定样本图像的最优哈希码。在3个通用数据集上进行了广泛的实验,结果有效表明了所提出有监督相似性保持的深度二阶哈希方法的有效性。

Recently,deep hashing technology,with its advantages of high storage efficiency and quick query speed,has been widely investigated in the field of large-scale visual image retrieval.However,the fundamental image features obtained by current deep hashing methods mainly depend on the first-order statistics of deep convolutional features,and they seldom take the global structure into consideration,leading to the limitation of retrieval accuracy to a certain degree.Focusing on the global representation capability and the similarity of intra-class samples,this paper proposes a novel supervised similarity preserving deep second-order hashing(S2PDSoH)based on deep pairwise supervised hashing,gaining effective performance improvement in the image retrieval task.Based on the pair-wise deep hashing model,S2PDSoH first employs covariance estimation based on matrix power normalization method to capture the deep second-order information of sample images,so that hash codes can possess good global presentation ability.Then,to gain more robust hash codes,it further constructs a joint constraint of category supervision and similarity preservation motivated by the idea of multi-loss integration,followed by an alternate iteration optimization algorithm to realize the end-to-end training.Therefore,with the semantic information added to the dual channel deep second-order hashing framework,S2PDSoH establishes a mechanism for common constraints on category supervision and similarity maintenance.In addition,it also introduces a hash-like function to achieve the approximate binarization result of hash codes,which solves the problem of non-convexity and avoids the quantization error in the hash mapping process.Extensive experimental results on three commonly used data sets show the effectiveness of the proposed deep-order hashing method with supervised similarity preservation.