Fast cross-modal retrieval method and system for incremental data carrying new categories

Abstract

The present invention proposes a fast cross-modal retrieval method and system for incremental data carrying new categories, including: an incremental hash learning step: extracting known category labels from known hash codes stored in a multimedia known category database Then, according to the similarity relationship between the existing category label and the unknown category label, the binary representation of the unknown category label is obtained, which is used to supervise the generation of the hash code of the incremental data in the incremental category database; the hash function Learning steps: In the learning process of the hash function, the anchor point set is obtained by sampling from the known category database and the incremental category database, based on the anchor point set, an asymmetric strategy is used to update the parameters of the deep network, and the hash function is Learn to get the desired model. The hash code of unknown incremental data can be directly learned while keeping the hash code of the original data unchanged, so as to meet the mode requirements of fast training.

Description

technical field [0001] The invention belongs to the technical fields of cross-modal retrieval and deep learning, and in particular relates to a fast cross-modal retrieval method and system for incremental data carrying new categories. Background technique [0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art. [0003] With the development of the Internet, multimedia data such as images, texts, videos, and audios generated by various smart terminal devices and websites show an explosive growth trend. Faced with these rich and massive multimedia data, how to quickly and accurately retrieve multi-modal data in the huge database according to user needs is a hot spot in the research of multimedia information retrieval. Therefore, cross-modal retrieval emerges as the times require, and its main purpose is to use query data of one modality to retrieve similar data of another modali...

AI Technical Summary

Problems solved by technology

[0005] However, the existing deep cross-modal hashing methods cannot solve the problem of incremental learning well

In other words, most deep cross-modal hashing methods cannot adapt to the incremental data of emerging unknown category labels in time, and need to be fed to the network at the same time as the original data to retrain the model

This mode has the following disadvantages: First, retraining all data means completely ignoring the hash code of the original data obtained through previous training, wasting computing resources and training time; in addition, as more and more unknown categories The generation of incremental data of tags, feeding the original data and incremental data to the network at the same time will increase the computational complexity of the model, which violates the demand for fast and accurate retrieval of large-scale multimedia data; and the current incremental hash learning Method cannot handle cross-modal retrieval issues

Images