A Multi-View Generation Method Based on Contrastive Learning

Abstract

In order to overcome the defect of lack of views in the appearance patent image collection, the present invention proposes a multi-view generation method based on contrastive learning, which includes the following steps: acquiring multi-view image data and corresponding real viewpoint labels; performing multi-view image data Preprocessing, constructing a training set; using contrastive learning constraints to train the encoder; connecting the decoder and the discriminator after completing the training to form a generative confrontation network, and inputting the training set into the generative confrontation network for confrontation training ; Input the appearance image, after the trained encoder extracts the view-invariant features, input the view-invariant features and the target view label into the trained decoder, the output is to retain the subject's intrinsic information and convert the view to the target view appearance image.

Description

technical field [0001] The present invention relates to the technical field of artificial intelligence, and more specifically, to a method for generating multi-views based on contrastive learning. Background technique [0002] With the advent of the era of knowledge economy, the intellectual property system plays an increasingly prominent role in the development and changes of society. As a protection object of intellectual property law, appearance design has gradually attracted people's attention. The number of design patent applications in my country is constantly increasing, and the number of applications has ranked first in the world. Facing the huge database of design patents, how to use computer technology to search design patents more efficiently has become an important research focus. At the same time, with The machine learning method represented by deep learning is one of the main research directions at present. The application of deep learning in computer vision ha...

AI Technical Summary

Problems solved by technology

At present, new view synthesis methods are divided into two categories, geometry-based methods and learning-based methods. However, they have the disadvantages of being limited to specific scenes and unable to infer 3D structural information, and the generated new perspective pictures will lose their original identity. Information, cannot effectively solve the problem of missing views in the appearance patent image set

Images