摘要：Faced with the huge image data in the context ofbig data era, how to effectively manage, describe, and retrievethem has become a hotspot issue in academia and industry. Inthis paper, we propose an end-to-end image retrieval systembased on deep convolutional neural network and differentiallearning method. Compared with the traditional method ofusing the deep convolutional activation features as the featurevector to match the image, we simplify the retrieval process ofthe method and decrease the problem of “semantic gap” in thecontent- based image retrieval system. We first build an imagematching dataset based on the gravitational field model, that isto add the similarity score label for each image in the datasetmanufacturing stage. Then we train the improved deep learningmodel and verify the effectiveness of the algorithm on threecommon image matching datasets (i.e., Caltech-101, Holidays,and Oxford Paris). Finally, the experimental results show thatour improved deep learning model with differential learningmethod that used for image retrieval system has state-of-the-artimage matching performance. The overall retrieval accuracy inCaltech-101, Holidays, and Oxford Paris datasets are 88.5%,94.1%, and 96.2%, respectively. As the number of returnedimage increases, the retrieval accuracy of the system decreasesslightly and eventually becomes stable at a high value. And thedifferential learning based retrieval method is superior to manytraditional algorithms in terms of image matching accuracy andsingle image processing speed.
关键词：image retrieval system; differential learning;deep convolutional neural network; gravitational field model;end-to-end training
