Cross-media training and retrieval method based on depth discrimination sorting learning

Abstract

The invention relates to a cross-media training and retrieval method based on depth discrimination sorting learning. The method comprises the following steps that: utilizing a deep network to extractfeatures from an image sample and a statement sample in a training set, and obtaining a feature vector pair, wherein the feature vector pair comprises an image feature vector used for showing the image sample, and a statement feature vector used for showing the statement sample; mapping the obtained feature vector pair to a common space, and calculating a similarity between the image feature vector and the statement feature vector; and utilizing a bidirectional discrimination sorting target function to sort the feature vector, and obtaining a training model.

Description

technical field [0001] The invention relates to the technical field of machine learning, in particular to a cross-media training and retrieval method based on deep discrimination and sorting learning. Background technique [0002] With the popularization and popularization of digital media technology, the amount of multimedia information with text, video, audio, graphics and images as the main body has increased significantly, and various new application requirements have also followed. As an important research direction in the field of multimedia and computer vision, cross-media retrieval has received extensive attention in recent years, and ranking learning algorithms have always been one of the important methods used in cross-media retrieval. [0003] In the prior art, there are a variety of ranking learning models that can be used for cross-media retrieval, such as passive-aggressive models, large-scale image annotation models, and supervised semantic indexing models. Si...

AI Technical Summary

Problems solved by technology

[0003] In the prior art, there are a variety of ranking learning models that can be used for cross-media retrieval, such as passive-aggressive models, large-scale image annotation models, and supervised semantic indexing models. Similar one-way ranking learning algorithms can be used to mine different media Semantic association between data, but ignores the structural information implicit in the queried modality; existing technologies also include sorting based on SVM (Support Vector Machine), WARP (Weighted Approximate Rank Pair-wise) or triples Models and other similar bidirectional ranking learning algorithms, such methods can embed semantic information into space, but cannot adjust feature representation according to specific tasks, and some of them do not consider the importance of samples (such as triplet sorting model) , some methods cannot effectively deal with large-scale data and high-dimensional data (such as SVM)

Images