Method and device for automatically adding and deleting portrait objects based on dichotomy labels

Abstract

The invention provides a method and device for automatically adding and deleting a portrait object based on a dichotomy tag, and the method comprises the steps: receiving an automatic adding and deleting request of the portrait object, and determining the type of at least one portrait object included in the automatic adding and deleting request of the portrait object; obtaining at least one face image to be modified and at least one reference face image; carrying out feature point detection on the face image to be modified and the reference face image; determining dichotomy labels of the to-be-modified face image and the reference face image for the type of portrait object through an addition and deletion module of the type of portrait object in a pre-trained neural network; and determining a target portrait object addition and deletion mode of the modified face image according to the portrait object features of the to-be-modified face image and the reference face image, and modifying the to-be-modified face image according to the target portrait object addition and deletion mode. According to the method provided by the invention, the corresponding portrait objects can be automatically added and deleted, the convenience is high, and the training cost is low.

Description

technical field [0001] The invention relates to the technical field of computer vision, in particular to a network method for adding and deleting portrait objects by using only binary classification labels. Background technique [0002] Portrait editing is an important application of computer vision. At present, people have a higher demand for the convenience of portrait photo editing. In related technologies, portrait editing is usually performed through an automatic method based on deep learning or a general image editing method. [0003] But most automated methods based on deep learning rely heavily on supervised data. For example, if you want to train a network that adds and deletes glasses, you may need to manually provide the semantic segmentation results of glasses as supervised data. Such supervised data requires a large amount of manual annotation, resulting in high training costs. [0004] Although general-purpose image editing methods (such as image editing sof...

AI Technical Summary

Problems solved by technology

These operations are not only cumbersome, but also often require users to have a high professional level to edit high-quality portraits

Images