具有性能感知排序的深度监督哈希用于多标签图像检索

现实生活中的图像大多具有多种标签属性。对于多标签图像，理想情况下检索到的图像应该按照与查询图像相似程度降序排列，即与查询图像共享的标签数量依次递减。然而，大多数哈希算法主要针对单标签图像检索而设计的，而且现有用于多标签图像检索的深度监督哈希算法忽略了哈希码的排序性能且没有充分地利用标签类别信息。针对此问题，提出了一种具有性能感知排序的深度监督哈希方法（deep supervised hashing with performance-aware ranking，PRDH），它能够有效地感知和优化模型的性能，改善多标签图像检索的效果。在哈希学习部分，设计了一种排序优化损失函数，以改善哈希码的排序性能；同时，还加入了一种空间划分损失函数，将具有不同数量的共享标签的图像划分到相应的汉明空间中；为了充分地利用标签信息，还鲜明地提出将预测标签用于检索阶段的汉明距离计算，并设计了一种用于多标签分类的损失函数，以实现对汉明距离排序的监督与优化。在三个多标签基准数据集上进行的大量检索实验结果表明，PRDH的各项评估指标均优于现有先进的深度哈希方法。

Deep supervised hashing with performance-aware ranking for multi-label image retrieval

Most images in real life have multi-label attributes. For multi-label images, ideally, the retrieved images should be ranked in descending order of similarity to the query image, namely their numbers of labels shared with the query image decrease sequentially. However, most hashing algorithms are mainly designed for the single label image retrieval, and the existing deep supervised hashing algorithms for multi-label image retrieval ignore the ranking performance of hash codes and do not fully utilize the label category information. To solve this problem, this paper proposed a deep supervised hashing with performance-aware ranking method(PRDH), which could effectively perceive and optimize the performance of the model and improve the effect of the multi-label image retrieval. In the hash learning part, this paper designed a ranking optimization loss function to improve the ranking performance of hash codes. At the same time, this paper adopted a spatial partition loss function to divide images with different numbers of shared labels into corresponding Hamming spaces. In order to fully utilize label information, this paper also explicitly proposed using predictive label for Hamming distance calculation in the retrieval stage, and designed a loss function for multi-label classification to achieve supervision and optimization of Hamming distance ranking. A large number of results of the retrieval experiments conducted in three multi-label benchmark datasets show that the evaluation metrics of PRDH outperform the state-of-the-art hashing approaches.