Picture generation method and device based on noise reduction stream, electronic equipment and storage medium

CN119228672APending Publication Date: 2024-12-31PING AN TECH (SHENZHEN) CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411389118.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-30
Publication Date
2024-12-31

AI Technical Summary

Technical Problem

The existing generative adversarial network (GAN) model has pattern collapse problems during training, and the model is low interpretability, resulting in a decrease in the quality of the generated results. Especially when expanding the model size or changing the data set, it requires multiple parameter adjustments to be stable. , affects the quality of image generation.

Method used

A picture generation method based on noise reduction stream is proposed. The image samples are noise-added and lost calculations are performed through the stream structure and the noise reduction autoencoder, and the target model is adjusted to generate the target model, and the high-quality samples are randomly sampled from the target generation model. picture.

Benefits of technology

It improves the training stability and generation quality of the image generation model, avoids training failures, and has a simple model structure and high interpretation, which is suitable for expansion into a large generative framework.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119228672A_ABST
    Figure CN119228672A_ABST
Patent Text Reader

Abstract

The embodiment of the invention provides a picture generation method and device based on a noise reduction stream, electronic equipment and a storage medium, and belongs to the technical field of artificial intelligence. The method comprises the following steps: carrying out noise addition processing on an obtained picture sample to obtain a noisy picture sample; performing first loss calculation on the noisy picture sample through the stream structure to obtain a first loss value; performing second loss calculation on the noisy picture sample through the noise reduction auto-encoder to obtain a second loss value; performing parameter adjustment on the picture generation model according to the first loss value and the second loss value to obtain a target generation model; random sampling is carried out on a first variable vector generated by training of the target generation model, and a second variable vector is obtained; and decoding the second variable vector through the target generation model to obtain a target picture. According to the embodiment of the invention, the second variable vector is input into the target generation model for decoding, so that the high-quality target picture can be generated, and the purpose of improving the picture generation quality is achieved.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Model training method and device and image generation method and device

    CN116630465A

  • Standardized stream-based audio generation method, apparatus and device, and storage medium

    CN117711416A

  • Method and device for generating image from text, storage medium and electronic equipment

    CN118587303A

  • Character recognition method and apparatus, device, medium, and product

    WO2023078070A1