Image editing method based on single-instance guided object representation splitting

Abstract

An image editing algorithm based on single-instance guided object representation splitting, including the following steps: 1) Construction of a supervision module based on a single example; 2) Dual strategy construction in a self-supervised module guided by a single example; 3) Single-sample Construction of Fuzzy Classification Strategies in Example-Guided Self-Supervision Module 4) Image Editing Based on Single-Example Annotated Image-Guided Object Representation Segmentation. The image editing algorithm based on single-sample-guided object representation splitting established through the above steps only needs to label one sample for each category of image to form a single-sample sample, and use the single-sample sample to guide a large number of unlabeled data training methods, through The single-sample supervision module and the single-sample-guided self-supervision module of unlabeled data realize the splitting of foreground objects and background representations in complex scenes, making it possible to directly manipulate images in the image representation space and easily realize related image editing tasks.

Description

technical field [0001] The invention belongs to the field of image representation learning (splitting), and can map the foreground object and background in the image to a representation with modular physical meaning. This representation can also be split into foreground object information and background information, which can be used in The high-dimensional representation space of images directly operates on foreground objects and backgrounds, and can be flexibly used in applications such as image editing. Aiming at the problems that supervised image representation splitting based on deep learning requires a large amount of labeled data and existing unsupervised methods are ineffective on complex background images, an image editing algorithm based on single-sample guided object representation splitting is proposed. Each Class image only uses one labeled sample (single sample) to guide all other unlabeled samples to learn, so that for a large amount of unlabeled data, a very sm...

AI Technical Summary

Problems solved by technology

[0006] The present invention aims to solve the problem that the current image object representation split based on deep learning requires a large number of labels and can only be applied to simple data, and provides an image editing algorithm based on single-sample guided object representation split

[0007] At present, image editing, such as exchanging the foreground objects of two images, can only be carried out in the image, which often requires Photoshop or the use of foreground and background segmentation models. If the size, position, and shape of the objects in the images are inconsistent, pixel black holes will appear. The problem of repairing, the current representation decoupling learning work can only realize the decoupling of attributes, such as color, size, etc., and it is often uncontrollable. Even if it is transferred to object representation splitting, a large number of labeled samples are required

In order to solve the problem that the deep network requires a large number of annotations and can only be applied to simple data, the present invention designs an image editing algorithm based on single-sample guided object representation splitting

Images