In the recent years, several hashing methods have been proposed for multi-label image retrieval. However, general methods quantify the similarities of image pairs roughly, which only consider the similarities based on category labels. In addition, general pairwise loss functions are not sensitive to the relative order of similar images. To address above problems, we present a deep semantic-aware ranking preserving hashing (DSRPH) method. First, we design a semantic-aware similarity quantization method which can measure fine-grained semantic-level similarity beyond the category based on the cosine similarity of image captions that contain high-level semantic description. Second, we propose a novel weighted pairwise loss function by adding adaptive upper and lower bounds, which can construct a compact zero-loss interval to directly constrain the relative order of similar images. Extensive experiments show that our method can generate high-quality hash codes and yield the state-of-the-art performance.

近年来，已经提出了几种用于多标签图像检索的哈希方法。但是，常规方法仅对图像对的相似度进行大致量化，而这些相似度仅基于类别标签来考虑。另外，一般的成对损失函数对相似图像的相对顺序不敏感。为了解决上述问题，我们提出了一种深层的语义感知排序保留哈希（DSRPH）方法。首先，我们设计了一种语义感知的相似度量化方法，该方法可以根据包含高级语义描述的图像标题的余弦相似度，测量超出类别的细粒度语义级相似度。其次，我们通过添加自适应上限和下限，提出了一种新颖的加权成对损失函数，它可以构造一个紧凑的零丢失间隔来直接约束相似图像的相对顺序。大量的实验表明，我们的方法可以生成高质量的哈希码，并产生最先进的性能。