Transfer Retrieval Method Based on Semi-Supervised Adversarial Generative Networks

Abstract

A migration retrieval method based on a semi-supervised confrontational generative network, by designing a confrontational generative network for hash retrieval across data domains, its goal is to map the original data set and the target data set to a common Hamming space, so that in a The image retrieval in a specific scene can be migrated to the retrieval image of another scene through the learning of the semi-supervised confrontation generation network, so as to solve the problems that unlabeled data cannot be fully utilized and the retrieval model is only suitable for a single scene in the era of big data. The invention effectively improves the automation and intelligence level of image retrieval.

Description

technical field [0001] The invention relates to the application of computer vision, pattern recognition, confrontation generation network, migration retrieval, cycle consistency, deep self-encoding convolutional neural network and deep learning technology in the field of image retrieval, in particular to a migration based on semi-supervised confrontation generation network search method. Background technique [0002] In the era of web 2.0, massive images, texts, and audio data are generated every day. How to quickly and accurately query the images that users need or are interested in in these vast and unlabeled data has become a research hotspot in the field of multimedia information retrieval. With the continuous development of artificial intelligence, the number of layers of convolutional neural networks continues to deepen. Whether image data, text data, or audio data can be trained, it can learn from a large amount of labeled data accurately from input to output. mappin...

AI Technical Summary

Problems solved by technology

[0005] In order to overcome the shortcomings of the existing image retrieval technology that needs to rely on a large amount of manually labeled data, the existing retrieval model is only suitable for a single scene, and does not effectively utilize massive unlabeled data, the present invention proposes a semi-supervised confrontation generation network based on The migration retrieval method, by designing a novel adversarial generation network for cross-data pre-hash retrieval, such as data under different databases, cameras or different scenarios, its goal is to map the original data set and the target data set to a Common Hamming space; use cyclic consistent network to maintain the similarity between similar images, and add edge hyperparameters to reduce the distance between similar images and increase the distance between dissimilar images; thus realizing cross-data domain The intelligent and fast image retrieval meets the image retrieval requirements in the era of big data

Images