Stable controllable image generation model training method based on W distance

An image generation and model training technology, applied in the field of image recognition, can solve the problems of unstable model training gradient and unstable model gradient descent direction, and achieve the effect of ensuring normal convergence, alleviating mode collapse, and increasing stability.

CN112070209AActive Publication Date: 2020-12-11HEBEI UNIVERSITY
4 Cites 12 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Publication Date
2020-12-11

Smart Images

  • Figure 1
    Figure 1
  • Figure 2
    Figure 2
  • Figure 3
    Figure 3
Patent Text Reader

Abstract

The invention provides a stable controllable image generation model training method based on a W distance, and the method comprises the following steps: a, preprocessing image data to acquire sample data of a training set; b, constructing a stable controllable image generation model based on a W distance; c, constructing an overall model according to a loss function of each network; d, alternatelyiterating model training by a gradient descent algorithm to ensure normal convergence of model parameters; and e, after the model parameters converge, respectively separating a generator network G, an encoder network E and a classifier network C as independent service products. The invention is an end-to-end network model, improves a model in the prior art from two aspects of a model structure and a distribution measurement standard, solves the problems of unstable model training gradient and unstable model gradient descending direction, can stably and directionally generate an image sample.Robustness in the model training process and controllability of the generated sample are improved.
Need to check novelty before this filing date? Find Prior Art

Description

technical field

[0001] The invention relates to the technical field of image recognition, in particular to a stable and controllable image generation model training method based on W distance. Background technique

[0002] In recent years, Generative Adversarial Networks (GAN) and its derived generative models have been one of the core topics in the machine learning and deep learning communities, among which BiGAN is the most important in high-dimensional complex data modeling. One of the deep generative models. The BiGAN model introduces the encoder network E into the GAN model so that the latent variables of similar samples can be gathered together during the encoding process, making the low-dimensional manifold continuous and achieving the effect of implicit regularization, which can improve the model's general ability.

[0003] However, in the training process of the BiGAN model, it is very easy for the support set of the real data distribution and the generated sample...

Examples

Embodiment Construction

[0078] The stable and controllable image generation (CWBLI) model training method based on W (Wasserstein) distance of the present invention comprises the following steps:

[0079] a. Preprocessing the image data to obtain the sample data of the training set and the test set.

[0080] Using the CelebA face dataset as sample data, divide the CelebA face dataset into a training set and a test set. Specifically, 180,000 images can be selected as the training set, and the remaining 22,599 images can be used as the test set. In the original CelebA dataset, each image has 178×218 pixels. In order to highlight the features we need and reduce the complexity of model training, each image in the training set and test set is cropped, such as Figure 11 As shown, the most classic 000001.jpg is taken as an example to show the processing flow, crop out a picture with a fixed face position in each image with a size of 64×64 pixels, and then normalize the cropped picture data.

[0081] b. Co...