Image representations based on deep learning models can provide exciting performance for image retrieval, but only using deep learning models cannot exploit global topological properties appropriately. The topological perception theory claims that the visual perception process is from global to local: global topological perception occurs earlier than other local patterns. Simulating the visual perception mechanism together with deep learning models to provide a compact yet discriminative representation remains challenging. Toward this end, we propose a novel image representation method called deep-embedding global feature descriptor. The main highlights include: (1) A frequency statistics ranking method is proposed to yield global topology features by combining global visual features and deep convolutional features. (2) An embedding method is proposed to embed the global topology feature spatially and channel-wise into deep convolutional features. It can reasonably integrate the global topological characteristic with local patterns by simulating the visual perceptual process from global to local. (3) A compact yet discriminative representation is provided by leveraging the advantages of global visual and deep features. Exhaustive experiments on five well-known benchmark datasets show that the proposed method outperforms some recent unsupervised state-of-the-art methods.

基于深度学习模型的图像表示可以为图像检索提供令人兴奋的性能，但仅使用深度学习模型无法适当地利用全局拓扑特性。拓扑知觉理论认为视觉知觉过程是从全局到局部的：全局拓扑知觉比其他局部模式发生得更早。将视觉感知机制与深度学习模型一起模拟以提供紧凑但有区别的表示仍然具有挑战性。为此，我们提出了一种新颖的图像表示方法，称为深度嵌入全局特征描述符。主要亮点包括：（1）提出了一种频率统计排序方法，通过结合全局视觉特征和深度卷积特征来产生全局拓扑特征。 (2)提出了一种嵌入方法，将全局拓扑特征在空间和通道上嵌入到深度卷积特征中。它通过模拟从全局到局部的视觉感知过程，能够将全局拓扑特征与局​​部模式合理地结合起来。 （3）利用全局视觉和深度特征的优势，提供紧凑而有区别的表示。对五个著名基准数据集的详尽实验表明，所提出的方法优于一些最近的无监督最先进方法。