Label-to-label-based multi-attribute prediction method and device, equipment and medium

Abstract

The invention relates to the technical field of artificial intelligence computer vision, in particular to a tag-to-tag-based multi-attribute prediction method and device, equipment and a medium, and the method comprises the steps: extracting depth features of image data; obtaining a pseudo tag sequence of the picture data based on the depth features; and masking each word in the pseudo tag sequence according to a preset probability to obtain a shielded sentence, recovering the shielded sentence to obtain a complete sentence, learning an association relationship among a plurality of attributes, and generating a final multi-attribute prediction result. Therefore, the problems that in a multi-task learning framework, the association relationship among multiple attribute label values cannot be modeled in one visual sample, and consequently the performance is poor are solved, the association relationship among multiple attributes is modeled by predicting the multiple shielded attribute label values, the performance of multi-attribute prediction is improved, and the accuracy of multi-attribute prediction is improved. Therefore, the accuracy of the model is effectively improved.

Description

technical field [0001] The present application relates to the technical field of artificial intelligence computer vision, and in particular to a label-to-label based multi-attribute prediction method, device, equipment and medium. Background technique [0002] The multi-attribute prediction problem in computer vision requires to predict multiple attribute values ​​for a given piece of visual data. [0003] In related technologies, common multi-attribute prediction applications include face attribute prediction, pedestrian attribute prediction, and clothing attribute prediction. Taking face attribute prediction as an example, for a given face picture, it is required to predict various attribute values ​​of the face at the same time, such as gender, whether to wear glasses, whether to have a beard, whether to be blond, whether to wear lipstick, etc. . Most of the existing methods use the framework of multi-task learning, that is, each attribute is regarded as a learning task...

AI Technical Summary

Problems solved by technology

[0005] This application provides a label-to-label based multi-attribute prediction method, device, equipment and medium to solve the problem that in the framework of multi-task learning, it is impossible to model the relationship between multiple attribute label values ​​in a visual sample, resulting in Poor performance and other issues

Images