We develop an image similarity descriptor for an image pair, based on deep features. The development consists of two parts - selecting the deep layer whose features are to be included in the descriptor, and a representation of the similarity between the images in the pair. The selection of the deep layer follows a sparse representation of the feature maps followed by multi-output support vector regression. The similarity representation is based on a novel correlation between the histograms of the feature maps of the two images. Experiments to demonstrate the effectiveness of the proposed descriptor are carried out on four applications that can be cast as classification tasks.

我们基于深层特征为图像对开发图像相似性描述符。该开发包括两个部分-选择要包含在描述符中的特征的深层，以及表示该对图像之间相似性的表示。深层的选择遵循特征图的稀疏表示，然后是多输出支持向量回归。相似性表示基于两个图像的特征图的直方图之间的新颖相关性。在四个可以转换为分类任务的应用程序上进行了实验，以证明所提出的描述符的有效性。