Cross-media multi-view imperfect label learning method

Abstract

The embodiment of the invention discloses a cross-media multi-view imperfect label learning method. The method comprises the following steps: preprocessing an imperfect label training set, and extracting two groups of features of the training set from two view angles which are conditionally independent; training the two groups of the features and the original imperfect labels in the training set in a new multi-label two-view flexible support vector machine, so as to obtain a group of training parameters; reclassifying the training set which is independent in the two view angles according to the two groups of the training parameters, so as to obtain two groups of independent classification results of the training set in the two view angles; randomizing the two groups of the classification results to obtain two groups of independent randomized classification results of the training set in the two view angles; acquiring new imperfect labels for the training set through a group of new supplementing and denoising algorithms; and terminating the iterative process until the change between the obtained new imperfect labels for the training set and the original imperfect labels is smaller than the set threshold value.

Description

technical field [0001] The invention belongs to the technical field of labels, and in particular relates to a cross-media and multi-view imperfect label learning method. Background technique [0002] With the advent of the information age, multimedia data has achieved explosive growth. Tags, as one of the content forms of multimedia, can help solve many important practical applications in data mining, especially in the field of cross-media, which plays a very important role. For example, using appropriate tags as part of image annotation, powerful image annotation and image retrieval techniques can be developed; using appropriate tags as part of movie reviews, an effective movie recommendation system can be developed; using appropriate tags as part of web page markup Part of it, a more efficient search engine can be developed. [0003] However, due to the rapid and explosive growth of data volume, it is unrealistic to rely solely on data processors to manually label all th...

AI Technical Summary

Problems solved by technology

It is precisely based on this characteristic that the results of social labeling are often imperfect, because every ordinary user who participates in social labeling cannot rule out his own subjectivity, carelessness, or even lack of patience to provide a perfect label

The imperfection of social labels can generally be reflected in two sub-problems: 1. Incomplete labels; 2. Noise labels

Incomplete labeling means that the given labels are correct, but cannot completely describe all the details of the digital resources, that is, some objects in the digital resources are missing labels

Images