Abstract:  Aiming at a low accuracy of image retrieval in a large number of images, this paper proposes a two-stage image retrieval model NL-VG which combines significance detection and convolution neural network. Firstly, in the first stage of the model, a nonlocal depth feature (NLDF) model combined with the local feature map and the global feature map is used to detect the saliency. Secondly, VGG-16 convolutional neural network is used to extract feature vectors in the second stage, which are then matched with the established image retrieval database through the similarity measurement method and similar images are displayed. Finally, the interactive interface toolkit PyQt5 is used to design the image retrieval system to realize the retrieval task. In this paper, web crawler technology is used to obtain and preprocess images to construct data sets. All images on the data sets are detected by the two-stage saliency detection model to obtain the image feature database. The experimental results show that the map value of the retrieval algorithm proposed in this paper is 0.767, which is more accurate than that of SpoC and other algorithms, and the query results are more consistent with the query expectations.