Handwritten numeral generation method based on parameter optimization generative adversarial network

Abstract

The invention provides a handwritten numeral generation method based on a parameter optimization generative adversarial network. The method comprises the following steps of preparing a handwritten digital data set as a sample training data set, sampling to obtain the real data, and initializing the random noise data; establishing the generative adversarial network, and initializing a generator network weight parameter theta and a discriminator network weight parameter omega; establishing a generator loss function and a discriminator loss function through the soil moving distance W, and additionally adding a gradient penalty loss item to the discriminator loss function; and iteratively training a generator network and a discriminator network, and optimizing the generator network weight parameter theta and the discriminator network weight parameter omega. According to the embodiment of the invention, the problems of slow convergence, unstable training, high calculation overhead and the like of the original generative adversarial network are solved, the optimization of the generative adversarial network is realized, the network performance of the generative adversarial network is fully improved, and the generator can generate the handwritten digital images with higher quality.

Description

technical field [0001] The embodiment of the present invention relates to a deep learning neural network, in particular to a method for generating handwritten digits based on parameter optimization generating an adversarial network. Background technique [0002] Generative Adversarial Network (GAN) is a method of training models proposed in 2014. This method refers to the minimal The idea of ​​​​a very large problem will eventually improve the effect of the two models. The goal of generating an adversarial network, given a set of real sample distributions, iteratively trains the generator network G and the discriminator network D according to the set, and finally enables the generator network G to generate samples that conform to the real sample distribution as much as possible from the noise signal , and the discriminator network D can judge from the distribution of the sample whether the sample conforms to the distribution of the real sample. [0003] The Lipschitz conti...

AI Technical Summary

Problems solved by technology

[0004] With the continuous development of generative confrontation network, it is widely used in various fields of artificial intelligence, but the original generative confrontation network is prone to problems such as gradient disappearance, unstable training, and slow convergence speed during the training process.

Since the original generation confrontation network uses JS divergence (Jensen-Shannon divergence) to measure the distance between generated samples and real samples, but in the training process, JS divergence is used to train to the end, regardless of the output quality of the generator, JS divergence is always is a fixed value, the loss of the discriminator cannot be used to measure the quality of the generated samples, and the output judgment of the discriminator is of little significance

[0005] Arjovsky Martin, Chintala Soumith, Bottou Leon.Wasserstein generative adversarial networks[C] / / International Conference on Machine Learning.2017:214-223. (Martin Ayovsky, Sumish Kintala, Leon Barton, Wasserstein Generative Adversarial Networks, International Conference on Machine Learning, 2017: 214-223), proposed an optimization method for generative adversarial networks, which has the following deficiencies: Gradient shearing method is used to realize the Lipschitz continuity condition , that is, to constrain all the parameters of the discriminator within a threshold range, which will easily cause the parameters of the discriminator to concentrate on the maximum value of the threshold range. In addition, this method has high requirements for setting the threshold, and setting the threshold too small will easily lead to gradient disappearance. If the threshold is set too large, it will easily lead to gradient explosion, resulting in poor optimization effect and difficult optimization

[0006] The Chinese invention patent document (CN 108470196A) disclosed on August 31, 2018 "A Method for Generating Handwritten Digits Based on a Deep Convolutional Adversarial Network Model" uses a deep convolutional adversarial network model to realize handwritten digit generation. This method has the following deficiencies : The generator loss function and the discriminator loss function of the generative adversarial network are not optimized separately, and the performance of the generative adversarial network cannot be fully utilized; the weight parameters of the generator network and the discriminator network are not optimized, which may easily lead to The discriminator network converges quickly, and loses its influence on the generator network prematurely, and the quality of the generated handwritten digital image is not high

Images